# Designing Agentive Technology > [!summary] Recommendation > Read Chapters 5–10 closely and Chapter 12 critically. Use the book as a pattern catalog for delegation, monitoring, correction, handoff, and evaluation. Do not use it as a current account of AI architecture, security, law, or empirical consensus. Its interaction model has aged better than its claims that smarter narrow AI becomes safer. > [!warning] Source boundary > The supplied EPUB is Layer 1. A targeted Readwise and Reader search found no exact title/author match, so no reader highlights, notes, colors, or emphasis were available. Every passage selected here is labeled `system-selected`. The EPUB is evidence about Noessel's text, not evidence about the reader's purpose or beliefs. **Attribution key:** **Author says** = explicit text · **Author implies** = reconstruction · **Evidence shows** = cited external evidence · **We infer** = system analysis · **We conclude** = current judgment · **Still uncertain** = source limit. ## Prompt 0 — Retrieval and vault setup ### Layer 1 ingestion summary | Check | Result | |---|---| | Source | Local EPUB 3 | | Integrity | `unzip -t` passed with zero compressed-data errors | | Identity | *Designing Agentive Technology: AI That Works for People*, Christopher Noessel, 2017 | | Publisher | Copyright page: Rosenfeld Media, LLC; OPF field: Lightning Source Inc. | | Identifiers | ISBN-10 1-933820-63-2; ISBN-13 978-1-933820-63-7; LCCN 2017934296 | | Edition | Not stated; no edition inferred | | Spine | 31 declared, 31 extracted in OPF order | | Text coverage | 2,289 addressable blocks; 61,549 whitespace-delimited words | | Chapter coverage | Introduction, 13/13 numbered chapters, two appendices, index, acknowledgments, author bio, and 19 endnotes read | | Images | 82 JPEG assets; 81 referenced; one unreferenced thumbnail | | Visual inspection | Appendix A's chronological touchpoint map and see-think-do map inspected | | Readwise | Targeted title/author searches returned no exact matching record or highlight | | Reader emphasis | Unavailable; 36 source excerpts below are `system-selected` | Stable anchors use `SP{spine number}:{filename}:B{block number}`. Example: `SP12:ch02.xhtml:B0062`. They identify exact blocks in this EPUB rather than pretending EPUB locations are universal page citations. ### Coverage map | Section | Purpose and central topic | Coverage | Representative source | |---|---|---|---| | Front matter and Introduction | Audience, structure, limited attention, and the case for systems that do work for people | Read | `SP05:fm01.xhtml:B0003–B0016`; `SP09:ch00.xhtml:B0003–B0010` | | Part I — Seeing | Defines agentive technology, argues for its value, and places it in automation history | Read | `SP10:part01.xhtml:B0003–B0005` | | Chapter 1 — The Thermostat That Evolved | Tool-to-agent story through temperature control | Read | `SP11:ch01.xhtml:B0010–B0031` | | Chapter 2 — Fait Accompli: Agentive Tech Is Here | Working definition and category boundaries | Read | `SP12:ch02.xhtml:B0038–B0109` | | Chapter 3 — Agentive Tech Can Change the World | Claimed benefits, scale, infrastructure, and future effects | Read | `SP13:ch03.xhtml:B0019–B0112` | | Chapter 4 — Six Takeaways | Myth, cybernetics, automation, feedback, and fluid agency | Read | `SP14:ch04.xhtml:B0010–B0104` | | Part II — Doing | Lifecycle patterns for designing and evaluating delegated work | Read | `SP15:part02.xhtml:B0003–B0004` | | Chapter 5 — A Modified Frame for Interaction | Agentive see-think-do and lifecycle touchpoints | Read | `SP16:ch05.xhtml:B0013–B0094` | | Chapter 6 — Ramping Up with an Agent | Capability, limits, goals, permissions, test drive, launch | Read | `SP17:ch06.xhtml:B0010–B0066` | | Chapter 7 — Everything Running Smoothly | Pause, monitoring, play, notifications, value | Read | `SP18:ch07.xhtml:B0010–B0049` | | Chapter 8 — Handling Exceptions | Trust, resources, false results, tuning, disengagement | Read | `SP19:ch08.xhtml:B0013–B0122` | | Chapter 9 — Handoff and Takeback | Vigilance, skill decay, practice, staged control transfer | Read | `SP20:ch09.xhtml:B0010–B0068` | | Chapter 10 — Evaluating Agents | Results, confidence, cooperation, perceived and measured value | Read | `SP21:ch10.xhtml:B0010–B0055` | | Part III — Thinking | Practice evolution, ethics, social effects, and call to action | Read | `SP22:part03.xhtml:B0003` | | Chapter 11 — How Will Our Practice Evolve? | Selling the concept, improving intelligence, fading agents | Read | `SP23:ch11.xhtml:B0007–B0031` | | Chapter 12 — Utopia, Dystopia, and Cat Videos | Bias, security, inequality, dependency, work, AGI | Read | `SP24:ch12.xhtml:B0015–B0132` | | Chapter 13 — Your Mission | Community of practice and human role | Read | `SP25:ch13.xhtml:B0007–B0018` | | Appendices and back matter | Consolidated touchpoints, 32-example catalog, index, 19 endnotes | Read; two key diagrams inspected | `SP26:app01.xhtml:B0001–B0004`; `SP27:app02.xhtml:B0002–B0074`; `SP31:footnote.xhtml:B0002–B0028` | ### Highlight index and cluster map There are no reader highlights to de-duplicate. The following source selections are `system-selected`, not reader emphasis. | Cluster | Location | Main topics | Rationale | Representative selections | |---|---|---|---|---| | A — Definition | Chapters 1–4 | agent, attention, delegation, automation, feedback | Establishes the category and its load-bearing distinctions | S-001–S-009 | | B — Lifecycle design | Chapters 5–8 | setup, visibility, permission, tuning, failure | Supplies the book's most reusable pattern language | S-010–S-024 | | C — Control and evaluation | Chapters 9–10 | vigilance, expertise, handoff, auditing, value | Tests whether delegated work remains governable | S-025–S-031 | | D — Practice and consequences | Chapters 11–13 | fading, bias, inequality, dependency, community | Extends product design into social responsibility | S-032–S-036 | ### System-selected source index | ID | Short passage | Anchor | |---|---|---| | S-001 | “They’ve begun to do things for you.” | `SP09:ch00.xhtml:B0008` | | S-002 | “an agent is a piece of narrow artificial intelligence” | `SP12:ch02.xhtml:B0040` | | S-003 | “It handles tasks so that you can use your limited attention” | `SP12:ch02.xhtml:B0060` | | S-004 | “a persistent, background assistant” | `SP12:ch02.xhtml:B0062` | | S-005 | differences warrant distinct planning and design | `SP12:ch02.xhtml:B0066` | | S-006 | agentive technology is “in service to a human” | `SP12:ch02.xhtml:B0099` | | S-007 | interests rather than moments | `SP13:ch03.xhtml:B0021` | | S-008 | goal-focused design | `SP13:ch03.xhtml:B0074` | | S-009 | “Think of agency as fluid.” | `SP14:ch04.xhtml:B0093` | | S-010 | the system does the heavy lifting | `SP16:ch05.xhtml:B0021` | | S-011 | routine agent reasoning is a black box to users | `SP16:ch05.xhtml:B0044` | | S-012 | tuning is constrained programming | `SP16:ch05.xhtml:B0046` | | S-013 | “The personal can be inferred from the impersonal.” | `SP16:ch05.xhtml:B0066` | | S-014 | convey what the agent cannot do | `SP17:ch06.xhtml:B0024` | | S-015 | minimize setup | `SP17:ch06.xhtml:B0031` | | S-016 | request private data only when necessary | `SP17:ch06.xhtml:B0035` | | S-017 | users need a “hood to look under” | `SP17:ch06.xhtml:B0048` | | S-018 | don't hide from users who check in | `SP18:ch07.xhtml:B0015` | | S-019 | support playing alongside the agent | `SP18:ch07.xhtml:B0019` | | S-020 | suggestions can undermine the point of delegation | `SP18:ch07.xhtml:B0027` | | S-021 | design agents to fail well | `SP19:ch08.xhtml:B0016` | | S-022 | trust builds slowly and falls quickly | `SP19:ch08.xhtml:B0026` | | S-023 | preview rule effects before saving | `SP19:ch08.xhtml:B0064` | | S-024 | allow graceful opt-out | `SP19:ch08.xhtml:B0114` | | S-025 | automation leaves people less equipped to take over | `SP20:ch09.xhtml:B0015` | | S-026 | vigilance decays during passive monitoring | `SP20:ch09.xhtml:B0019` | | S-027 | staged warnings preserve context | `SP20:ch09.xhtml:B0045–B0046` | | S-028 | familiar persistent maps beat hidden context | `SP20:ch09.xhtml:B0050` | | S-029 | external auditing must verify the agent | `SP21:ch10.xhtml:B0028` | | S-030 | compare measured outcomes with a non-agent baseline | `SP21:ch10.xhtml:B0046–B0048` | | S-031 | evaluation spans results, confidence, cooperation, and value | `SP21:ch10.xhtml:B0049–B0055` | | S-032 | agents that fade can improve humans | `SP23:ch11.xhtml:B0024` | | S-033 | “No technology is neutral.” | `SP24:ch12.xhtml:B0015` | | S-034 | algorithms can encode authors' bias | `SP24:ch12.xhtml:B0037` | | S-035 | unequal agent access can harden social stratification | `SP24:ch12.xhtml:B0059–B0066` | | S-036 | build a shared practice, vocabulary, cases, and metrics | `SP25:ch13.xhtml:B0009–B0013` | ## Prompt 1 — Setup, contract, baselines, biases, and priors ### Reading contract > [!caution] 🔶 Inferred reader inputs > Purpose, existing knowledge, constraints, and DEIB goals weren't supplied. The working purpose below is a system inference from the selected title and the Work the Book request. Correct it before treating it as the reader's own contract. - **Why read:** 🔶 Assess whether a 2017 interaction-design framework remains useful for current AI-agent product and workflow design. - **Hope to understand:** The lifecycle touchpoints that keep delegated work understandable, correctable, interruptible, and measurable. - **Practical use:** Apply the strongest patterns as review questions for agentic products and personal automations. - **Success criteria:** Recover the full argument; preserve source anchors; distinguish useful patterns from unsupported forecasts; identify what current standards add. - **Evidence that would change the analysis:** Comparative evaluations showing that the lifecycle patterns don't improve outcomes, control, or recovery; or current evidence that a criticized claim holds across the scope stated. - **What the book cannot establish:** Current product capability, present law, population-level social effects, causal effectiveness of its pattern catalog, or the reader's priorities. - **Constraints:** One consolidated note; EPUB-only Layer 1; no exact Readwise match; no adjacent note or Reader mutations; 2017 snapshot; sparse bibliography; raster accessibility defects. ### Prior beliefs and incentives | Area | Reader state | System-analysis safeguard | |---|---|---| | Prior belief about agents | 🔶 Not supplied | Don't convert current familiarity with AI agents into a reader claim. | | Existing HCI knowledge | 🔶 Not supplied | Define terms in the note; don't assume expertise. | | Desired conclusion | 🔶 Not supplied | Preserve both confirmation and disconfirmation. | | Professional incentive | 🔶 Not supplied | Note that product-design relevance could bias the review toward actionable patterns. | | Technology optimism or fear | 🔶 Not supplied | Grade forecasts separately from design guidance. | | Financial or ideological incentive | 🔶 Not supplied | No inference. | ### Open questions inventory **Definitions and mechanisms** 1. Does “acting on behalf” describe behavior, governance, intent, or marketing? 2. Can agentive and assistive modes coexist within one task without a clean boundary? 3. Is persistence necessary, or can a short autonomous run still be agentive? 4. Does the author's automation distinction survive the automation literature's broader definitions? **Evidence and credibility** 5. Which patterns rest on human-factors research rather than product anecdotes? 6. What comparative evidence shows that an agent reduces workload rather than shifting it to supervision? 7. Are user confidence, trust calibration, and measured performance separated well enough? 8. How should correlated examples from one author's practice affect confidence? **Application and failure** 9. What permissions, scopes, and revocation controls make “on behalf” enforceable? 10. Which failure modes require prevention rather than graceful recovery? 11. What makes a handoff timely when the human has lost context or skill? 12. How should an agent expose false negatives that neither agent nor user can directly observe? 13. What baseline establishes that the agent adds value? 14. When should a team choose assistance or deterministic automation instead? **History and current context** 15. How do LLM-based tool-using agents change the book's narrow-AI assumptions? 16. Which 2017 forecasts have become ordinary infrastructure, and which remain speculative? 17. How do current risk frameworks extend the book's product-level ethics? **DEIB and accessibility** 18. Who can afford the hardware, data, connectivity, time, and expertise assumed by the examples? 19. Do users affected by an agent have standing when they aren't its purchaser or operator? 20. How should disability-led design change the book's separation of assistive and agentive technology? 21. Can a visual pattern language be taught from an EPUB whose diagrams lack descriptive alt text? **Unanswered or avoided** 22. Who is accountable when user goals, provider incentives, and affected-party rights conflict? 23. What logs and appeal mechanisms should exist for high-impact decisions? 24. How do multi-agent coordination and adversarial inputs alter the lifecycle? ### Priors seed table These are system-analysis priors, not reader beliefs. | Claim | Initial belief | Prior confidence | Reason | Evidence needed | |---|---|---:|---|---| | C-001 | A persistent delegated-work category is useful | 75% | Fits common product behavior | Boundary tests | | C-005 | Agentive design is distinct from automation | 45% | Useful intent lens; contested taxonomy | Automation literature | | C-007 | Limits and permissions belong in onboarding | 85% | Strong risk and HAI precedent | Comparative usability evidence | | C-010 | Trust must track performance | 85% | Established human-factors concern | Domain-specific studies | | C-011 | Automation can weaken takeover readiness | 85% | Long human-factors lineage | Replications and scope | | C-013 | External audit is necessary | 90% | Self-evaluation lacks an independent oracle | Evaluation studies | | C-016 | Agents free attention | 60% | Plausible, but supervision may offset gains | Workload measurements | | C-018 | Agents can deepen inequality | 70% | Access and service-tier mechanisms are credible | Distributional studies | | C-021 | Agent rules can seed safe AGI ethics | 10% | Speculative leap | Mechanism and validation | | C-022 | Smarter narrow AI becomes safer | 30% | Accuracy can rise while attack surface grows | Risk by capability and access | | C-023 | Agents should fade as competence grows | 60% | Fits scaffolding in learning contexts | Transfer and relapse studies | | C-024 | The field needs shared patterns and metrics | 85% | Professional practice has moved this way | Adoption and outcome evidence | ### Drift assumptions - **Fast drift:** model capabilities, agent tool access, benchmarks, security attacks, provider products, regulation, and vocabulary. Recheck at least annually. - **Moderate drift:** common HAI guidelines, identity and authorization patterns, deployment practice, labor evidence. - **Low drift:** vigilance limits, skill decay risk, the need for baselines, and the distinction between confidence and measured performance. - **Fixed:** what Noessel argued in this 2017 text. ## Prompt 2 — Inspectional reading ### Book map and central project The author explicitly structures the book as a logical argument (`SP05:fm01.xhtml:B0009–B0012`): Part I asks readers to see a category, Part II gives a pattern language for designing it, and Part III considers practice and consequences. | Part | Apparent purpose | Relationship to argument | Evidence available | Confidence | |---|---|---|---|---:| | Seeing, Chapters 1–4 | Define and justify agentive technology | Supplies the category and premises | Product histories, named research, examples | High on map; moderate on proof | | Doing, Chapters 5–10 | Give lifecycle patterns and evaluation heuristics | Carries the book's main practical value | Pattern catalog, thought experiment, practitioner cases | High | | Thinking, Chapters 11–13 | Project practice changes and ethical effects | Adds cautions and a call to community | Ethical analysis and forecasts | High on content; low-to-moderate on forecasts | - **Main problem:** Traditional tool-centered interaction design doesn't cover systems that keep working while the user's attention is elsewhere. - **Proposed answer:** Treat persistent narrow-AI delegation as its own design category and design the full service lifecycle. - **Method:** Conceptual classification, historical examples, practitioner patterns, one extended thought experiment, and ethical speculation. - **Audience:** Product owners, technology strategists, interaction designers, educators, students, futurists, and science-fiction creators (`SP05:fm01.xhtml:B0003–B0007`). - **Rhetorical modes:** explain, persuade, teach, advocate, warn, and speculate. ### Provisional summary *Provisional after inspectional reading:* Noessel argues that systems which persistently monitor conditions and act for a person's goals deserve treatment as a distinct interaction-design category. The category matters because the person no longer performs every task; they set direction, supervise selectively, correct exceptions, receive control during failures, and judge whether delegation produced value. The book first defines this shift through tools, automation history, and narrow AI. It then offers lifecycle patterns for setup, permission, monitoring, notifications, tuning, handoff, takeback, disengagement, and evaluation. Its practical center is a call to design delegated work as an ongoing relationship rather than a collection of screens. The final section recognizes security, bias, unequal access, dependency, ecological cost, labor disruption, and skill loss, but often frames them as questions rather than governed requirements. The result is a strong interaction-pattern catalog with uneven evidence and forecasts that need current verification. ### Reading hotspots - **Chapter 2:** Definition and contested boundaries; the thesis depends on them. - **Chapters 5–8:** Core lifecycle and exception patterns; highest practical value. - **Chapter 9:** The strongest research-grounded risk: vigilance, expertise, and control transfer. - **Chapter 10:** Baselines and independent evaluation; crucial for validating value. - **Chapter 12:** Bias, equity, safety, dependency, and labor; broadest claims and highest drift. - **Appendix A:** Two visual summaries of the pattern system; the EPUB's poor alt text makes direct inspection necessary. ### Guiding questions 1. Is the category boundary necessary for the patterns to work? 2. Which touchpoints are universal, risk-tiered, or obsolete? 3. Does “user goal” include affected parties and provider incentives? 4. Can the user meaningfully review a system whose reasoning and context are opaque? 5. How should teams test false negatives and long-horizon failures? 6. When does monitoring recreate the work that delegation promised to remove? 7. Which handoff patterns work when users lack expertise? 8. Does a fake-money test drive predict real financial behavior? 9. Are perceived value and trust being mistaken for correctness? 10. How should access control, audit logs, rollback, and redress enter the lifecycle? 11. Which claims still fit LLM-based agents? 12. What changes when agents coordinate with other agents? 13. Whose interests disappear when the purchaser is treated as “the user”? 14. What would falsify the claim that agentive framing improves design outcomes? ### Reading strategy **Recommendation: selective deep read.** Read Chapters 5–10 and 12 in full; use Chapters 1–4 to understand the author's vocabulary; skim the product catalog and dated forecasts. Change this recommendation to a syntopical comparison when using the book for high-stakes system design, pairing it with human-factors research, current HAI guidance, risk management, security, accessibility, and law. ## Prompt 3 — Classification and terminology engine ### Classification - **Domain:** Human-computer interaction and product design. - **Subdomain:** Human-AI interaction, automation, delegated-work UX. - **Genre:** Practitioner framework and pattern catalog with speculative ethics. - **Argument:** Definitional, historical, practical, evaluative, and prescriptive. - **Primary method:** Examples plus conceptual synthesis, not systematic review. - **Intended use:** Frame agentive product strategy, scenarios, touchpoints, and evaluation. - **Comparison traditions:** Cybernetics; automation and function allocation; mixed-initiative interaction; calm technology; human-centered AI; service design. ### Controlled vocabulary | Term | Author's definition or use | Competing definition | Operational definition here | Equivocation risk | Example / counterexample | Evidence | |---|---|---|---|---|---|---| | Agentive technology | Narrow AI acting for a user in the background | Current “agentic AI” often means LLM planning and tool use | Persistent delegated work with monitored conditions and external action | High | Google Alert / calculator | `SP12:ch02.xhtml:B0040–B0062` | | ANI | Intelligence restricted to a domain | Modern foundation models cross many domains | System with bounded task competence | Medium | Spam filter / general assistant | `SP12:ch02.xhtml:B0042–B0056` | | AGI | Human-like general problem solving | No settled operational threshold | Broad cross-domain capability with transferable planning | High | Speculative / thermostat | `SP12:ch02.xhtml:B0045–B0048` | | Tool | Stays within the user's active task | Tools may include autonomous features | User directly drives each task step | Medium | Light switch / scheduled thermostat | `SP12:ch02.xhtml:B0058–B0060` | | Assistive | Helps the person remain primary task-doer | Disability usage names technologies that improve access | System advises or augments while human acts | High | Navigation cue / background autopilot | `SP12:ch02.xhtml:B0082–B0084` | | Automation | Author: aims to remove the human | Field usage includes varied human-machine allocation | Rule- or model-driven execution with predefined control allocation | High | Assembly line / co-created drafting | `SP12:ch02.xhtml:B0098–B0099` | | Persistence | Work continues beyond a direct interaction | Long-running process needn't imply agency | Monitors or acts after the initiating moment | Low | Saved search / one-shot lookup | `SP12:ch02.xhtml:B0060–B0062` | | Trigger | Condition for action | Learned policies may not expose a discrete trigger | Auditable condition that initiates a behavior | Medium | temperature threshold / free-form goal only | `SP17:ch06.xhtml:B0037–B0041` | | Exception | Condition defeating the general rule | Some systems treat exceptions as model uncertainty | Known case requiring suppression, change, or escalation | Medium | skip this sender / unobserved failure | `SP19:ch08.xhtml:B0036–B0080` | | Behavior | What the agent does | Current agent behavior may emerge from planning | Authorized external action and method | Medium | send alert / infer without acting | `SP19:ch08.xhtml:B0081–B0090` | | Constraint | Bound on action or resources | A prompt rule may be advisory, not enforced | Machine-enforced scope, budget, actor, or time limit | High | spending cap / prose request | `SP17:ch06.xhtml:B0041` | | Touchpoint | Human entry into delegated work | Service design uses touchpoint more broadly | Moment of understanding, choice, correction, transfer, or exit | Low | launch / hidden background step | `SP16:ch05.xhtml:B0021–B0043` | | Trust | Confidence built through interactions | Calibrated trust should track warranted reliance | Match between reliance and measured capability | High | justified delegation / confident misuse | `SP19:ch08.xhtml:B0019–B0027` | | Handoff | Agent transfers task/control | May include escalation to another system or operator | Explicit transfer with context, authority, readiness, and acceptance | Medium | staged takeover / abrupt failure | `SP20:ch09.xhtml:B0028–B0054` | | Takeback | Control returns to agent | Could be partial or shared | Confirmed resumption after stability review | Medium | manual agreement / silent seizure | `SP20:ch09.xhtml:B0055–B0062` | | Measured value | Objective outcome difference | “Objective” still depends on metric choice | Predefined outcome against a relevant baseline | Medium | error-rate delta / satisfaction alone | `SP21:ch10.xhtml:B0044–B0048` | | Perceived value | User's cost-benefit judgment | Adoption can reflect lock-in or lack of choice | Reported worth plus observed continued use and alternatives | High | voluntary retention / forced use | `SP21:ch10.xhtml:B0039–B0040` | | Fluid agency | Agency varies by task and context | Autonomy may also be modeled by authority and risk | Per-capability allocation of sensing, deciding, and acting | Low | assistive during exception / one global level | `SP14:ch04.xhtml:B0077–B0093` | ### Contested language - **Agent, agentive, agentic:** The author's user-facing category and current backend architecture aren't synonyms. Noessel later called them “cousins,” distinguishing UX from infrastructure. - **On behalf of:** Can hide provider incentives, affected parties, or unenforced authority. - **Smart, intelligence, understands:** Risk anthropomorphism and capability inflation. - **Trust and confidence:** Shouldn't stand in for reliability, rights, or accountability. - **Automation:** The author's intentional distinction is narrower than common human-factors usage. - **Slave and personhood:** The ethical metaphor can distract from present human labor, ownership, and power. ### Category-error watchlist | Risk | Status | Finding | |---|---|---| | Correlation vs causation | Present | Product examples show coexistence, not that agentive framing caused value. | | Individual vs institutional behavior | Present | “User goal” can omit provider incentives and organizational control. | | Description vs prescription | Present | Category emergence is used to support a required design shift. | | Intention vs effect | Present | Being “in service” doesn't prove beneficial or rights-respecting outcomes. | | Anecdote vs population evidence | Present | Most examples lack denominators or comparative outcomes. | | Technical capability vs practical outcome | Present | Capability lists don't prove reliable service. | | Absence of evidence vs evidence of absence | Guarded | The review doesn't treat uncited harms or missing counterexamples as nonexistent. | | Legal vs moral claims | Present in Chapter 12 | Ethics and liability questions are exploratory, not legal guidance. | | Theological vs historical claims | Not materially applicable | Religious references are rhetorical rather than load-bearing theology. | ## Prompt 4 — Claims inventory, dependency graph, and evidence expectations ### Claims inventory `EA` = explicit author claim; `IA` = implied author claim. No reader-inference claim is attributed to the author. | ID | Claim and source | Type / scope / source | Testability and expected evidence | Non-vacuous falsifier | Key terms / chapter | |---|---|---|---|---|---| | C-001 | Agentive technology is persistent narrow AI acting for a user's goal while attention is elsewhere. `SP12:ch02.xhtml:B0040–B0062` | Definitional; broad; EA | Boundary-test varied systems | Counterexamples classified consistently without persistence or user-goal service | agentive, persistence; Ch. 2 | | C-002 | Technologies often evolve from manual tools through powered, metrical, assistive, and agentive forms. `SP12:ch02.xhtml:B0027–B0038` | Historical; broad; EA | Cross-domain historical sample | Many mature domains that don't follow the sequence | tool, assistive; Ch. 2 | | C-003 | Moving monitoring out of attention creates user value. `SP12:ch02.xhtml:B0058–B0062` | Causal/value; task-level; EA | Workload and outcome studies | Supervision and correction cost equals or exceeds saved attention | attention; Ch. 2 | | C-004 | Assistive systems help a person do a task; agentive systems perform it for them. `SP12:ch02.xhtml:B0082–B0084` | Taxonomic; EA | Mixed-mode case analysis | Systems routinely switch modes without useful boundary | assistive; Ch. 2 | | C-005 | Agentive design differs from automation because it remains explicitly in human service. `SP12:ch02.xhtml:B0098–B0099` | Taxonomic/normative; EA | Compare literature and incentives | Automation systems meet the same service criterion or agent systems optimize provider goals | automation; Ch. 2 | | C-006 | The category warrants distinct planning, design, and evaluation. `SP12:ch02.xhtml:B0064–B0079` | Prescriptive; practice; EA | Comparative project outcomes | Ordinary service/HCI methods cover all material needs equally well | touchpoint; Ch. 2 | | C-007 | Setup must convey capability and limits, collect goals and permissions, support testing, then launch. `SP17:ch06.xhtml:B0010–B0058` | Prescriptive; lifecycle; EA | Onboarding and incident studies | Removing a step improves comprehension and safety without new harm | permission, test drive; Ch. 6 | | C-008 | Users need monitoring and proportionate notifications without forced vigilance. `SP18:ch07.xhtml:B0010–B0049` | Prescriptive; EA | Attention, notification, detection studies | Quiet or continuous-monitoring designs outperform across equivalent risk | monitoring; Ch. 7 | | C-009 | False positives and false negatives require distinct correction paths. `SP19:ch08.xhtml:B0036–B0080` | Methodological; EA | Classification-task usability | One correction method handles both with equal discovery and control | trigger, exception; Ch. 8 | | C-010 | Trust rises through success and can collapse after serious failure. `SP19:ch08.xhtml:B0019–B0027` | Behavioral; EA | Longitudinal calibrated-trust data | Trust changes symmetrically or independently of failure severity | trust; Ch. 8 | | C-011 | Handoff is a weak point because passive automation reduces vigilance, context, and expertise. `SP20:ch09.xhtml:B0010–B0027` | Causal; safety-critical; EA | Human-factors experiments and incidents | No takeover loss after controlling for workload, training, and reliability | handoff; Ch. 9 | | C-012 | Practice, staged warnings, persistent maps, and intermediaries can improve takeover. `SP20:ch09.xhtml:B0028–B0054` | Prescriptive; risk-tiered; EA | Transfer and incident measures | These patterns delay or worsen takeover versus alternatives | practice; Ch. 9 | | C-013 | Agent correctness needs an external audit because the agent can't independently validate its own model. `SP21:ch10.xhtml:B0025–B0029` | Methodological; EA | Independent-oracle comparison | Self-checking matches independent audit across known failure classes | audit; Ch. 10 | | C-014 | Evaluation must separate results, user confidence, cooperation, perceived value, and measured value. `SP21:ch10.xhtml:B0030–B0055` | Evaluation taxonomy; EA | Construct validity and use in field studies | Measures collapse without lost diagnostic value | baseline, confidence; Ch. 10 | | C-015 | Users should be able to play alongside, direct, or take control without breaking goals. `SP18:ch07.xhtml:B0017–B0020`; `SP21:ch10.xhtml:B0041–B0043` | Prescriptive; EA | Directability and task studies | Direct control creates more errors or burden than benefit in the domain | play, cooperation; Chs. 7, 10 | | C-016 | Agents can reduce burden and extend access to useful services. `SP13:ch03.xhtml:B0036–B0077` | Causal/value; broad; EA | Distributional outcome studies | Net burden or exclusion rises in representative use | access, goal; Ch. 3 | | C-017 | Competing agents and infrastructure changes will follow wide adoption. `SP13:ch03.xhtml:B0080–B0106` | Predictive; societal; EA | Adoption, platform, and market evidence | Agent adoption remains isolated without coordination or infrastructure effects | infrastructure; Ch. 3 | | C-018 | Unequal access to agents and human fallback can deepen stratification. `SP24:ch12.xhtml:B0059–B0066` | Predictive/equity; EA | Service-tier and outcome data | Low-resource users receive equal capability, redress, and outcomes | equity; Ch. 12 | | C-019 | Delegation can cause skill loss and fragile dependence. `SP24:ch12.xhtml:B0079–B0093` | Causal; societal; EA | Longitudinal proficiency and resilience studies | Delegated users retain skill and recover during outages as well as active users | skill, dependence; Ch. 12 | | C-020 | Rapid automation can displace jobs faster than institutions retrain people. `SP24:ch12.xhtml:B0099–B0106` | Predictive/economic; EA | Labor transitions and task-level data | Displaced workers consistently transition without widening duration or loss | labor; Ch. 12 | | C-021 | Human-authored agent rules might supply ethics training for future AGI. `SP24:ch12.xhtml:B0128–B0130` | Speculative; IA | Mechanism, representativeness, alignment tests | Rules encode conflict and bias or fail to generalize safely | AGI, ethics; Ch. 12 | | C-022 | Narrow AI becomes safer as it becomes smarter. `SP06:fm02.xhtml:B0013` | Causal/safety; broad; EA | Risk tracked against capability and access | Capability adds attack paths or consequence without equivalent controls | safety, ANI; FAQ | | C-023 | Some agents should fade as users gain competence. `SP23:ch11.xhtml:B0019–B0024` | Prescriptive/learning; EA | Transfer and retention studies | Fading harms outcomes or competence doesn't transfer | scaffold; Ch. 11 | | C-024 | A community needs shared vocabulary, cases, patterns, libraries, and metrics. `SP25:ch13.xhtml:B0009–B0013` | Prescriptive/professional; EA | Practice adoption and outcome quality | Shared artifacts don't improve coordination or results | practice; Ch. 13 | | C-025 | Agentive framing is a productive design lens more than a discrete testable property. `SP12:ch02.xhtml:B0100–B0104` | Methodological; EA | Design comparison | The lens adds no distinct decisions or produces worse ones | framing; Ch. 2 | ### Evidence expectations summary The book supplies named human-factors research for C-011, product and practitioner examples for many design claims, and logical argument for the taxonomy. It rarely supplies controlled comparisons, representative samples, denominators, or independent replication. The empirical burden is therefore strongest where the book predicts universal value, safety, social impact, or labor effects. ### Claim dependency graph ```mermaid flowchart TD C001["C-001: persistent delegated-work category"] --> C006["C-006: distinct design practice"] C002["C-002: tool-to-agent trajectory"] -. weak historical bridge .-> C001 C003["C-003: attention value"] --> C016["C-016: reduced burden"] C004["C-004: assistive/agentive distinction"] --> C015["C-015: directable mode switching"] C006 --> LIFE["Lifecycle patterns C-007–C-015"] LIFE --> C024["C-024: shared practice"] C010["C-010: trust dynamics"] --> C011["C-011: handoff risk"] C011 --> C012["C-012: takeover mitigations"] C013["C-013: external audit"] --> C014["C-014: multidimensional evaluation"] C016 --> C017["C-017: infrastructure effects"] C017 --> C018["C-018: stratification risk"] C017 --> C019["C-019: dependence risk"] C022["C-022: smarter means safer"]:::weak --> FUT["Optimistic future case"] C021["C-021: rules seed AGI ethics"]:::weak --> FUT classDef weak stroke-dasharray: 5 5,stroke:#b33,color:#b33 ``` **Load-bearing:** C-001 and C-006 for the category; C-011 and C-013 for the safety/evaluation case. **Single-point weakness:** C-022 can't support the optimistic safety case as stated. The practical patterns survive its failure. ### Claims mapped to open questions - C-001–C-006 → Q1–Q6, Q15. - C-007–C-015 → Q7–Q14, Q21, Q24. - C-016–C-020 → Q17–Q20, Q22. - C-021–C-025 → Q14–Q17, Q23–Q24. ## Prompt 5 — Quantitative rigor setup ### Update policy 1. Treat confidence numbers as structured judgment, not measured probability. 2. Start with the claim's scope and base rate; don't reward a claim for rhetorical clarity. 3. Separate source quality, relevance, independence, and direction. 4. Count repeated examples from Noessel as one authorial evidence stream. 5. Give independent human-factors studies more weight than product anecdotes. 6. Exclude moral and definitional claims from numerical updating when data can't adjudicate them. 7. Name the single input driving each update and retain a qualitative fallback. ### Confidence bands | Label | Range | Meaning here | |---|---:|---| | Very low | 0–20% | Speculative or contradicted at stated scope | | Low | 21–40% | Plausible but weakly supported or poorly scoped | | Uncertain | 41–60% | Mixed evidence or unresolved boundary | | Moderate | 61–75% | Useful support with material limits | | High | 76–90% | Strong support, not universal proof | | Very high | 91–100% | Direct, independent, and unusually well-bounded | ### Priors and posteriors This table was defined at Prompt 5 and populated only after the Prompt 7 claim evaluation. Its placement preserves the prompt sequence; the updates don't precede the evidence review in the session record. | Claim | Prior | Evidence considered | Likelihood impact | Posterior | Band | Main reason | |---|---:|---|---|---:|---|---| | C-001 | 75 | Full category definition plus current author distinction | Moderate increase | 80 | High | The lens still maps common delegated systems. | | C-005 | 45 | Automation literature treats allocation more broadly | Moderate decrease | 35 | Low | Intent doesn't create a clean disciplinary boundary. | | C-007 | 85 | Microsoft HAI guidance, NIST, EU oversight duties | Moderate increase | 90 | High | Capability, limits, correction, and oversight recur independently. | | C-010 | 85 | Human-automation trust literature and current HAI guidance | Small increase | 88 | High | Trust calibration remains a live design problem. | | C-011 | 85 | Bainbridge and Parasuraman-Sheridan-Wickens | Moderate increase | 90 | High | Handoff and out-of-loop risks have independent support. | | C-013 | 90 | Current evaluation practice and independent-oracle logic | Small increase | 92 | Very high | Self-assessment cannot establish ground truth alone. | | C-016 | 60 | Examples show benefit; supervision and access costs remain | No meaningful update | 60 | Uncertain | Net burden is domain- and design-dependent. | | C-018 | 70 | Book's service-tier mechanism plus current rights/risk frameworks | Small increase | 75 | Moderate | Mechanism is credible; distributional outcomes remain undermeasured here. | | C-021 | 10 | No validated bridge from user rules to AGI alignment | Moderate decrease | 5 | Very low | Metaphor substitutes for a learning and governance mechanism. | | C-022 | 30 | Current agent reliability and hijacking evidence | Moderate decrease | 15 | Very low | Capability can add reach and attack surface. | | C-023 | 60 | Pedagogical scaffolding analogy but little source evidence | Small decrease | 55 | Uncertain | Applicable to learning agents, not a general agent principle. | | C-024 | 85 | HAI guidelines, risk frameworks, and current agent standards work | Small increase | 90 | High | The field did build shared guidance and evaluation practice. | ### Calibration warnings | Risk | Severity | Control | |---|---|---| | False precision | High | Read numbers as ordered judgments only. | | Double-counting | High | One book, one authorial evidence stream. | | Correlated examples | High | Product examples often share markets and assumptions. | | Survivorship bias | High | Failed or abandoned agents rarely enter the catalog. | | Publication bias | Moderate | Positive design cases are more visible than quiet failures. | | Base-rate neglect | Moderate | Risk depends on task frequency, impact, and user population. | | Tiny samples | High | Several claims rely on one project or anecdote. | | Selection effects | High | The 32 products were chosen for explanatory familiarity. | | Unavailable denominators | High | Few failure, adoption, or outcome rates appear. | | Proxy measures | High | Confidence and retention don't equal correctness or welfare. | | P-values as effect sizes | Not applicable | The book reports almost no inferential statistics. | ### Non-quantitative fallback Use **strongly decreases**, **moderately decreases**, **no meaningful update**, **moderately increases**, or **strongly increases** when a number would hide a normative judgment or evidence gap. This is the default for ethics, category boundaries, and speculative futures. ## Prompt 6 — Analytical reading scaffold ### Chapter analysis | Chapter | Purpose, thesis, and claims | Evidence, examples, and counterexamples | Assumptions, questions, and role in spine | Source IDs | |---|---|---|---|---| | 1 | Shows a thermostat lineage ending in a learning agent | Historical artifacts and Nest; no negative product history | Assumes one trajectory generalizes; opens the category case | `SP11:ch01.xhtml:B0010–B0031` | | 2 | Defines agentive as persistent narrow AI acting for a user | Writing, music, search, pet feeder; boundary cases acknowledged | Taxonomy is pragmatic, not settled; foundational | `SP12:ch02.xhtml:B0038–B0109` | | 3 | Claims agents shift attention from moments to interests and goals | Alerts, transport, finance, civic systems, speculative infrastructure | Benefits often outrun measured outcomes; expands stakes | `SP13:ch03.xhtml:B0019–B0112` | | 4 | Replaces simple function allocation with feedback and fluid agency | Fitts, Bainbridge, cybernetics, autonomy critiques | Historical synthesis is selective; repairs Part I's simplicity | `SP14:ch04.xhtml:B0010–B0104` | | 5 | Recasts see-think-do with system-led routine work | Diagrams, technology inventory, Mr. McGregor | Treats reasoning as black box from user view; organizes patterns | `SP16:ch05.xhtml:B0013–B0137` | | 6 | Makes setup a trust and authority transition | Capability, limitation, permission, test-drive patterns | Test drives may not reproduce live risk; begins lifecycle | `SP17:ch06.xhtml:B0010–B0105` | | 7 | Keeps routine running visible without attention overload | Monitoring and notification taxonomy | Notification efficacy and accessibility are untested; normal operations | `SP18:ch07.xhtml:B0010–B0070` | | 8 | Treats exceptions as primary design material | False-positive/negative tools, rule builder, opt-out | Assumes users can program and diagnose; failure management | `SP19:ch08.xhtml:B0013–B0138` | | 9 | Explains takeover as an out-of-loop problem | Vigilance, skill decay, Navy practice, staged alarms | Domain transfer from critical systems needs care; weakest operational link | `SP20:ch09.xhtml:B0010–B0068` | | 10 | Defines multidimensional agent evaluation | Prototypes, field study, audit, confidence, value baseline | Heuristics need validation; supplies the evidence gate | `SP21:ch10.xhtml:B0010–B0055` | | 11 | Predicts practice will sell, improve, then hide agents | Industry experience and forecasts | “Fade away” conflates calm UX and skill scaffolding; looks ahead | `SP23:ch11.xhtml:B0007–B0031` | | 12 | Names non-neutral effects and ethical questions | Bias, inequality, dependency, labor, AGI thought experiments | Breadth exceeds evidence; vital caution with speculative bridges | `SP24:ch12.xhtml:B0015–B0132` | | 13 | Calls for shared models, cases, libraries, and metrics | Science-fiction contrasts and practitioner appeal | Community doesn't guarantee accountability; conclusion | `SP25:ch13.xhtml:B0003–B0018` | ### Chapter-slot completion Reader responses weren't supplied for any chapter. The final column therefore records a `system-selected` response without presenting it as the reader's view. | Chapter | Key terms | Strongest counterexample or alternative | Unanswered question | Relation to argument / system response | |---|---|---|---|---| | 1 | tool, metrical, assistive, agentive | Many technologies remain useful tools rather than climbing a fixed ladder. | How representative is temperature control of other domains? | Opens the category case; useful history, weak general trajectory. | | 2 | ANI, persistence, on behalf, automation | Mixed-initiative systems cross the proposed boundaries during one task. | Which observable test separates agentive design from ordinary automation? | Foundational definition; keep as a lens rather than an ontology. | | 3 | attention, interests, goals, infrastructure | Supervision and recovery may consume more attention than the original work. | What net workload and distributional outcomes result? | Supplies the value premise; needs comparative evidence. | | 4 | cybernetics, feedback, fluid agency | Authority and risk can explain allocation better than one agency scale. | How should fluid agency be measured per capability? | Repairs the simple ladder; strengthens task-level mode switching. | | 5 | see-think-do, touchpoint, black box | Service design and incident response may cover the same lifecycle. | Which touchpoints are unique to agents? | Organizes the practical core; high reuse value. | | 6 | capability, limits, goal, permission, test drive | Simulated conditions may hide live data, stakes, or adversaries. | What makes consent informed and authority enforceable? | Starts the lifecycle; translate interface permission into system controls. | | 7 | monitoring, pause, play, notification | Quiet operation can conceal failures; frequent notices recreate the burden. | What interruption rate preserves detection without forced vigilance? | Defines normal operation; test by risk tier and access need. | | 8 | trust, trigger, exception, rule, disengagement | Users may lack the skill or evidence needed to diagnose a false negative. | What independent process discovers unseen misses? | Makes failure primary design material; pair correction with external audit. | | 9 | vigilance, expertise, handoff, takeback | Full manual control isn't always safer when context or skill has decayed. | Who accepts control when the operator isn't ready? | Establishes the strongest safety case; require staged, practiced transfer. | | 10 | results, confidence, cooperation, perceived value, measured value | Retention or satisfaction can stay high while outcomes or rights worsen. | Which baseline and affected-party metrics count? | Supplies the evidence gate; extend beyond purchaser value. | | 11 | persuasion, improvement, fading | Calm presentation and educational scaffolding solve different problems. | When is fading an explicit user goal? | Projects practice change; narrow fading to measured skill transfer. | | 12 | neutrality, bias, inequality, dependency, labor, AGI | Public provision and enforced rights can change the predicted distribution. | Who owns prevention, appeal, remedy, and compensation? | Raises necessary risks; lacks mechanisms and representative evidence. | | 13 | patterns, cases, libraries, metrics, community | Shared language can standardize bad assumptions or exclude affected groups. | Who governs the practice and revises its evidence? | Ends with collective practice; require plural participation and audit. | **Reader response:** 🔶 Not supplied. **System response:** Chapters 5–10 form a coherent lifecycle review tool; Chapters 3 and 12 need the strongest evidence and scope checks. ### Argument spine 1. People have scarce attention and recurring goals. 2. Some systems persist and act outside the immediate interaction. 3. **C-001:** This behavior forms a useful category called agentive technology. 4. **C-006:** Category differences create lifecycle touchpoints that ordinary screen-by-screen design can miss. 5. **C-007–C-015:** Design must cover setup, routine visibility, exception repair, transfer, exit, and evaluation. 6. **C-016–C-020:** Wide deployment changes access, skills, work, and infrastructure. 7. **C-024:** Practitioners need shared patterns and evidence. **Weakest link:** The move from a useful framing to broad claims about value and social benefit. Repair it by treating the category as a checklist generator, then requiring domain-specific baselines, rights analysis, threat modeling, and independent evaluation. ### Evidence ledger | Evidence ID | Claim IDs | Type and source | Independence | Quality | Limits | Source IDs | |---|---|---|---|---|---|---| | E-001 | C-001–C-006 | Author's cross-domain product synthesis | No | C | Selected cases; no comparison | `SP11–SP14` | | E-002 | C-011 | Bainbridge, “Ironies of Automation” (1983) | Yes | B | Critical industrial systems | `SP14:ch04.xhtml:B0047`; footnote 19–20 | | E-003 | C-011–C-012 | Mackworth vigilance work | Yes | B | Specific monitoring task | `SP20:ch09.xhtml:B0018–B0022` | | E-004 | C-012 | Don Norman automation examples | Partly | C | Talk/example, not field trial | `SP20:ch09.xhtml:B0044–B0054` | | E-005 | C-007–C-015 | Mr. McGregor thought experiment | No | D | Speculative and author-designed | `SP16–SP19` | | E-006 | C-013–C-014 | Evaluation logic and product metrics | No | C | Proposed heuristics lack validation | `SP21:ch10.xhtml:B0025–B0055` | | E-007 | C-007, C-010 | Microsoft 2019 HAI guideline validation | Yes | B | 49 practitioners, 20 products; guideline study | Current context | | E-008 | C-011 | Parasuraman, Sheridan, Wickens (2000) | Yes | B | Model and literature synthesis | Current comparison | | E-009 | C-007, C-013, C-018 | NIST AI RMF 1.0 | Yes | B | Voluntary framework, not outcome trial | Current context | | E-010 | C-007, C-012, C-018 | EU AI Act human oversight duties | Yes | B | Applies by system/risk/legal scope | Current context | | E-011 | C-022 | NIST agent hijacking evaluation | Yes | B | LLM agents, broader than fixed narrow AI | Current context | | E-012 | C-017, C-022 | International AI Safety Report 2026 | Yes | B | Synthesis with remaining evidence gaps | Current context | Independent evidence items: six. Independent evidence directly challenges C-022 and narrows C-016–C-020; it supports the handoff, oversight, and evaluation core. ### Reading plan and completion criteria - **Priority:** Chapters 5–10, then 12, then the definitional Chapters 2 and 4. - **Skim:** Dated product catalog and AGI forecasts unless researching history. - **External verification:** Safety, privacy, labor, law, current products, and agent reliability. - **Comparison terms:** automation, autonomy, agentic AI, trust calibration, human oversight. - **Synthesis stops:** after definition, lifecycle, evaluation, and social-risk sections. - **Attention:** High for Chapters 8–10 and 12; medium elsewhere. Prompt 7 may begin because the source is fully covered; claims, falsifiers, source labels, evidence ledger, and external context are present. Missing reader purpose limits personal prioritization, not claim evaluation. ## Prompt 7 — X-ray analysis and fair criticism ### Strongest reasonable reconstruction **Author says:** A persistent system that acts while the person's attention is elsewhere changes the person's role from task-doer to task-manager. **We infer:** The book's strongest argument doesn't require agentive technology to be a universally distinct ontology. It only requires delegation to create recurring lifecycle risks that ordinary interaction flows miss. **We conclude:** That repaired form is faithful to the practical center and survives the category's contested edges. ### Claim-by-claim evaluation | ID | Support | Grade | Strongest objection / counterexample | Current conclusion | |---|---|---:|---|---| | C-001 | Definition plus many cases | N/A | Current agents may be broad, short-lived, or provider-serving | Useful operational lens, not settled taxonomy | | C-002 | Five-domain historical stories | D | Technological change isn't a fixed ladder | Keep as heuristic, reject as general history | | C-003 | Attention rationale and examples | C | Oversight can recreate workload | Plausible only when net burden is measured | | C-004 | Clear mode contrast | N/A | Mixed-initiative systems switch continuously | Use as modes, not product types | | C-005 | Intent-based distinction | D | Automation can serve people; agents can serve firms | Not a reliable boundary | | C-006 | Lifecycle gaps are concrete | B | Service design could absorb the same needs | Distinct checklist is justified even if category isn't | | C-007 | Detailed setup patterns; later HAI support | B | Consent screens can be formalities | Strong when permissions are enforced and revocable | | C-008 | Notification taxonomy | C | Notifications burden users and miss accessibility needs | Risk-tier and test interruption cost | | C-009 | Classification logic | B | Users may never see false negatives | Strong, but needs external discovery and audit | | C-010 | Design observation and trust research | B | Trust varies by domain, culture, and consequence | Replace generic trust-building with calibration | | C-011 | Human-factors evidence | A- | Full automation or continuous assistance changes the case | Strongest empirical claim; scope by risk and allocation | | C-012 | Plausible mitigation set | B | Practice and alerts can fail under surprise | Use in layered recovery, not as guarantees | | C-013 | Independent-oracle logic | A- | Audit can share the same data or model failure | Strong if the oracle is genuinely independent | | C-014 | Separates key constructs | B | Omits rights, security, and affected-party outcomes | Retain and extend | | C-015 | Directability patterns | B | Manual control can be unsafe in some moments | Make directability risk- and state-dependent | | C-016 | Many benefit examples | C | Burden can shift to setup, correction, or excluded people | Context-dependent, not universal | | C-017 | Current agent infrastructure partly confirms | C+ | Infrastructure could centralize without agent competition | Directionally supported; mechanism varies | | C-018 | Clear price/service-tier mechanism | C+ | Public provision or standards can reduce gaps | Serious risk, not inevitable outcome | | C-019 | Skill-decay lineage and resilience logic | B | Some skills transfer or don't need maintenance | Strong for practiced, time-critical tasks | | C-020 | Historical analogy and forecast | C | Task change, new roles, policy, and adoption pace vary | Live question; book doesn't establish magnitude | | C-021 | Thought experiment | F | Human rules are biased, conflicting, and not a validated alignment corpus | Speculation only | | C-022 | Assertion in FAQ | F | Capability increases reach, attack surface, and consequence | False at stated scope; narrow sharply | | C-023 | Learning-scaffold analogy | C | Many agents aren't educational and shouldn't manipulate dependence | Useful only for explicit competence-building goals | | C-024 | Need stated; later field practice confirms | B | Shared vocabulary can standardize bad assumptions | Strong with plural participation and evidence | | C-025 | Author's own pragmatic framing | N/A | A lens can still mislead | Best interpretation of the category claim | **Grade distribution:** A-/A: 2 · B: 9 · C/C+: 7 · D: 2 · F: 2 · N/A: 3. The pattern shows a credible lifecycle and human-factors core, weak universal taxonomy, and unsupported speculative safety claims. ### Prior-update dependence check - C-001 and C-005 draw mainly from one authorial stream; Noessel's 2024 distinction clarifies his terms but isn't independent confirmation. - C-007 and C-024 gain independent convergence from Microsoft and NIST guidance, but guidance agreement isn't an outcome trial. - C-010 and C-011 draw from related human-factors traditions; they aren't counted as unrelated replications. - C-013 rests chiefly on independent-oracle logic; repeated reviewer examples don't raise its confidence. - C-016 and C-018 remain limited because product examples and governance frameworks don't supply representative workload or distributional outcomes. - C-021 receives no valid empirical bridge. C-022 receives a strong downward update from current agent-security evidence, with a scope warning that LLM agents are broader than the book's fixed narrow-AI examples. - C-023 relies on a learning-scaffold analogy rather than direct evidence from the book's agent cases. ### Biases, weaknesses, and counter-observations | Issue | Finding | Effect | |---|---|---| | Selection bias | Examples are chosen for explanatory fit and visibility. | Moderate; weakens prevalence and value claims. | | Survivorship bias | Failed products and rejected designs are sparse. | Moderate; inflates feasibility. | | Anecdotal generalization | Product and personal stories often stand in for comparative evidence. | Strong on broad causal claims. | | Equivocation | “Agent,” “automation,” “trust,” and “smart” shift between technical and normative senses. | Moderate. | | Appeal to novelty | “New category” helps motivate practice change. | Mild; practical patterns don't depend on novelty. | | Optimism bias | Future benefit often receives more mechanism than future harm. | Moderate; Chapter 12 partly corrects it. | | Unfalsifiable framing | AGI ethics corpus proposal lacks a testable bridge. | Strong for C-021 only. | Counter-observations in the author's favor: - He states that the definition is a productive lens rather than a discrete property (`SP12:ch02.xhtml:B0100–B0104`). - He treats failure as inevitable and makes exceptions central (`SP19:ch08.xhtml:B0013–B0016`). - He says an external audit must verify results (`SP21:ch10.xhtml:B0028`). - He recognizes bias, unequal service, ecology, dependency, labor, and non-neutrality in Chapter 12. - He concedes that the ethical inventory isn't complete (`SP24:ch12.xhtml:B0132`). ### Repairs | Claim | Repair | Faithful? | Cost | |---|---|---|---| | C-002 | Some domains show a tool-to-agent trajectory; don't claim a general law. | Yes | Loses evolutionary inevitability. | | C-005 | Agentive design emphasizes accountable service to a person; automation is a broader allocation field. | Yes | Gives up the clean boundary. | | C-010 | Design for warranted reliance, not maximum trust. | Yes | Replaces persuasive onboarding with ongoing evidence. | | C-016 | Agents reduce burden only when setup, oversight, recovery, and exclusion costs remain lower than displaced work. | Yes | Makes value conditional and measurable. | | C-018 | Without access and redress protections, agent tiers can deepen inequality. | Yes | Changes prediction to governed risk. | | C-022 | Within fixed scope and authority, better accuracy may reduce ordinary errors; added capability or access requires separate risk controls. | Partly | Abandons the slogan that smarter means safer. | | C-023 | Fade only when the user's explicit goal is skill acquisition and transfer is measured. | Yes | Narrows it to learning contexts. | | C-021 | Not repairable as evidence-backed AGI policy from this source. | No | Removing it doesn't harm the practical framework. | ### Attribution lanes - **Author says:** Claims C-001–C-025 as anchored. - **Evidence shows:** Independent research supports handoff risk, oversight, calibration, and external evaluation; current security evidence contradicts C-022's broad safety claim. - **We infer:** The lifecycle checklist is the book's durable core. - **We conclude:** Treat category boundaries and forecasts as hypotheses, not facts. - **Still uncertain:** How the patterns perform across cultures, disabilities, power relations, and modern LLM-agent architectures. ## Prompt 8 — Systems, context, and historiography ### Intellectual lineage ```mermaid flowchart LR Myth["Myths and automata"] --> Cyber["Cybernetics and feedback"] Turing["Turing-era computing"] --> HCI["Human-computer interaction"] Fitts["Fitts function allocation"] --> Irony["Bainbridge: ironies of automation"] Cyber --> Irony Irony --> Mixed["Mixed-initiative and human-machine teamwork"] HCI --> Mixed Mixed --> Noessel["Noessel 2017: agentive lifecycle"] Noessel --> HAI["2019 human-AI interaction guidelines"] Noessel --> Risk["2023–2026 risk, oversight, and agent security"] ``` The arrows from Noessel to later work show chronology and conceptual convergence, not documented causal influence. ### Context map | Context | Available to author in 2017 | Added after publication | Effect on reading | |---|---|---|---| | Technical | Narrow ML, sensors, recommenders, IoT, early autonomous systems | Foundation models, tool-using agents, long-horizon planning, multi-agent systems | “Narrow AI” no longer describes every product called an agent. | | HAI practice | Automation and mixed-initiative research | Microsoft's 18 HAI guidelines validated in 2019 | Independently supports limits, correction, explanation, and adaptation. | | Risk governance | Product ethics and scattered sector rules | NIST AI RMF, generative-AI profile, EU AI Act | Adds lifecycle governance, rights, documentation, oversight, and accountability. | | Security | Privacy inference, spoofing, classic system security | Indirect prompt injection, tool misuse, identity, delegated authority | “On behalf” now requires strong authorization and hostile-input design. | | Economic | Consumer automation and job forecasts | Wider generative/agent deployment and uneven labor evidence | Forecasts need current task- and sector-level data. | | Cultural | U.S.-weighted consumer-product frame | Broader design-justice and affected-community practice | Purchaser-centered “user goals” are insufficient. | In 2024 Noessel distinguished **agentive** front-end strategy and UX from **agentic** backend planning and tool orchestration. That later clarification preserves the book's interaction focus while acknowledging a technical category the 2017 text didn't cover. ### Incentives and missing perspectives - Product firms benefit when delegation becomes a differentiator, data source, subscription, or service-tier boundary. - Designers benefit professionally from a named category and pattern language; this doesn't invalidate it, but it explains the advocacy. - The record is shaped mainly by designers, technologists, consumer products, transport, finance, and classic automation research. - Missing or thin: disabled users as co-designers, low-connectivity communities, frontline workers displaced or supervised by agents, people subject to decisions without buying the product, non-U.S. governance, labor organizers, and environmental-justice perspectives. ## Prompt 9 — DEIB, multicultural, accessibility, and reading-level audit ### Inclusion strengths - Chapter 12 names unequal access, premium human fallback, encoded bias, ecological cost, dependency, and labor displacement. - The framework values user correction, opt-out, monitoring, and control rather than passive compliance. - Fluid agency avoids one universal allocation of human and machine work. - The book acknowledges its U.S.-specific time-use source and doesn't claim a complete ethics inventory. These are rhetorical and conceptual strengths. They don't amount to guaranteed representation, procedural protection, redress, or affected-community governance. ### Omission and exclusion risks - The central user is often a purchaser or operator with a home, smartphone, data access, garden, portfolio, vehicle, or paid service. - “Acting on behalf” centers one user's goal and underdevelops conflicts with workers, bystanders, communities, and public institutions. - Disability appears mainly in a boundary discussion about assistive technology, not as a source of design authority. - Race, gender, neurodiversity, language difference, and colonial or Indigenous data governance receive little sustained analysis. - The “99%” section recognizes class stratification but doesn't specify participation, appeal, or enforcement mechanisms. - Cultural transfer is uncertain when values, consent, household structure, infrastructure, and trust in institutions differ. ### Multicultural transferability | Pattern | Transferability | Risk | Adaptation | |---|---|---|---| | Capability and limit disclosure | High | Literacy and language mismatch | Plain language, localization, multimodal explanation, teach-back | | Goal and preference collection | Moderate | Individualist model of goals | Support household, collective, and fiduciary goals plus conflicts | | Permission | Moderate | Consent may be coerced or bundled | Granular scope, revocation, local norms, independent redress | | Notifications | Moderate | Different urgency, device, and attention norms | User-tested channels, quiet modes, accessible alternatives | | Rule tuning | Low-to-moderate | Requires technical and linguistic confidence | Examples, previews, undo, assisted setup, representative testing | | Handoff | High in principle | Expertise and authority differ by institution | Name responsible humans, escalation path, language access | | Measured value | Moderate | Metrics encode provider values | Co-define outcomes with affected groups; report distribution | ### Accessibility audit | Dimension | Finding | |---|---| | EPUB navigation | Strong TOC, page anchors, headings, and spine order. | | Images | Fails descriptive access: 79 substantive images use only `alt="Images"`; two covers use `cover-image`. | | Raster-only meaning | Appendix A's diagrams carry topology not fully present in prose. | | Language | Conversational and often explains jargon, but uses dense references, humor, acronyms, and long conceptual passages. | | Syntax | Approximate Flesch Reading Ease 52.9; Flesch–Kincaid grade 10.7 using a deterministic heuristic across prose blocks. | | Reference density | High: products, science fiction, automation scholarship, and historical figures can burden readers even when sentences are readable. | | Cognitive load | The many lists and lifecycle states need summaries, stable labels, and progressive disclosure. | | Practical access | Several proposed agents assume connected devices, private data, sensors, money, and ongoing service availability. | ### Supplementary perspectives - Sasha Costanza-Chock, *Design Justice*, for community-led design and power. - Safiya Umoja Noble, *Algorithms of Oppression*, and Ruha Benjamin, *Race After Technology*, for structural bias. - Aimi Hamraie, *Building Access*, and disability-led design practice for access as expertise. - Human-factors work on automation, function allocation, and out-of-loop performance. - Current HAI guidelines, NIST risk management, agent security work, and applicable law. ## Prompt 10 — Notes pipeline The one-file contract controls this run. These are note candidates inside the consolidated review; no separate permanent notes, literature notes, MOC edits, or reciprocal links were created. ### Fleeting notes - “On behalf of” needs a rights and authority test, not only a user-intent statement. - The lifecycle map resembles service design plus incident response more than screen design. - Monitoring is a tax; the right question is whether the tax is lower than the displaced work. - False negatives need sampling or an outside oracle because absence is invisible. - A test drive with fake stakes may build false confidence if it doesn't reproduce real conditions. - “Trust” should become “warranted reliance” in any operational checklist. - Directability conflicts with safety when a panicked user can override a better-performing system. - Disengagement should include data export, credential revocation, delegated-task cleanup, and successor ownership. - Multi-agent conflict turns “cooperation” into a protocol and governance problem. - Chapter 12 asks the right categories of questions but rarely assigns responsibility. ### Literature notes | Suggested title | Note | Source | |---|---|---| | Lit · Agentive technology shifts attention from tasks to delegated interests | The author's core value claim is relief from recurring monitoring, subject to net-burden measurement. | C-003; `SP12:ch02.xhtml:B0058–B0062` | | Lit · Agent setup is an authority transition | Capability, limits, goals, permissions, testing, and launch define what the system may do. | C-007; `SP17:ch06.xhtml:B0010–B0058` | | Lit · Exception design is the center of agent UX | People most need the interface when triggers, behaviors, resources, or context fail. | C-009; `SP19:ch08.xhtml:B0013–B0122` | | Lit · Handoff debt grows during successful automation | Reliable routine operation reduces vigilance, context, and practiced skill before failure. | C-011; `SP20:ch09.xhtml:B0010–B0027` | | Lit · A takeover requires context before control | Staged warnings, persistent maps, and readiness signals help rebuild the situation model. | C-012; `SP20:ch09.xhtml:B0044–B0062` | | Lit · An agent cannot be its only evaluator | Correctness needs an external audit and a relevant baseline. | C-013–C-014; `SP21:ch10.xhtml:B0025–B0048` | | Lit · Confidence and measured value are separate | A user can trust a poor agent or dislike an effective one; both constructs need evidence. | C-014; `SP21:ch10.xhtml:B0030–B0048` | | Lit · Fluid agency beats a single autonomy level | Allocation changes by task, context, risk, and capability. | C-004, C-015; `SP14:ch04.xhtml:B0077–B0093` | | Lit · Agent access can stratify service | Cheap rigid service for most and flexible human service for wealthy users can harden inequality. | C-018; `SP24:ch12.xhtml:B0059–B0066` | | Lit · No technology is neutral but questions are not controls | Ethical awareness must become ownership, enforcement, audit, and redress. | C-018–C-020; `SP24:ch12.xhtml:B0015–B0132` | ### Permanent-note candidates #### Agent authority needs a revocable contract, not a claim of good intent - **Suggested filename:** `Agent Authority Needs a Revocable Contract Not a Claim of Good Intent.md` - **Idea:** A system acts “on behalf” only when identity, scope, data access, action rights, duration, audit, and revocation are explicit and enforced. - **Confidence:** High. - **Scope and conditions:** Applies to systems that read private data, change state, spend resources, or communicate externally. A local read-only recommendation tool needs less ceremony. - **Supports:** C-001, C-005, C-007. - **Links:** [[3. Resources/MOCs/AI Agent Safety MOC]], [[3. Resources/Permanent Notes/Bounded Action Spaces Not Better Models Make Agents Safe to Run]], [[3. Resources/Permanent Notes/Delegated Credentials Must Attenuate Rather Than Escalate]]. #### Handoff debt accumulates while automation appears successful - **Suggested filename:** `Handoff Debt Accumulates While Automation Appears Successful.md` - **Idea:** Each successful delegated run can reduce the human's context and practice, raising the cost of the rare takeover. - **Confidence:** High in monitoring and practiced-skill domains. - **Scope and conditions:** Strongest for time-critical or complex tasks; weak for low-stakes tasks with easy recovery. - **Supports:** C-011–C-012. - **Links:** [[3. Resources/Permanent Notes/Automation Removes the Social Checks a Process Silently Depended On]], [[3. Resources/MOCs/AI Agent Safety MOC]]. #### An agent's value is the outcome delta after supervision cost - **Suggested filename:** `An Agent's Value Is the Outcome Delta After Supervision Cost.md` - **Idea:** Compare outcomes with a baseline after counting setup, monitoring, correction, recovery, and externalized harm. - **Confidence:** High. - **Scope and conditions:** Metric choice must include affected parties and distribution, not only purchaser efficiency. - **Supports:** C-003, C-014, C-016. - **Links:** [[3. Resources/Permanent Notes/An Agent Evaluation Without an Incumbent Baseline Cannot Establish Value]], [[3. Resources/Permanent Notes/AI ROI Requires a Pre-Existing Performance Baseline]]. #### False negatives require an observer outside the agent's model - **Suggested filename:** `False Negatives Require an Observer Outside the Agent's Model.md` - **Idea:** A system that defines its own ground truth will classify every detected positive as correct and cannot count missed events without an independent source. - **Confidence:** Very high. - **Scope and conditions:** The external observer must not share the same blind spot, data, or model assumption. - **Supports:** C-009, C-013. - **Links:** [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]], [[3. Resources/Permanent Notes/External Test Oracles Make AI-Generated Code Trustworthy]]. #### Capability growth and safety growth are separate curves - **Suggested filename:** `Capability Growth and Safety Growth Are Separate Curves.md` - **Idea:** Better task performance can reduce ordinary errors while broader tools, authority, and reachable systems add new failure and attack paths. - **Confidence:** High for tool-using agents. - **Scope and conditions:** A fixed-scope classifier can become safer as accuracy improves; don't transfer that result to expanded authority. - **Supports:** C-022. - **Links:** [[3. Resources/Permanent Notes/Capability and Intent Are Separate Axes of Risk]], [[3. Resources/Permanent Notes/A Site-Modifying Agent Needs Enforced Guardrails Before It Needs More Capability]]. #### Directability is a state-dependent safety property - **Suggested filename:** `Directability Is a State-Dependent Safety Property.md` - **Idea:** The right amount of human control changes with task state, expertise, consequence, and system confidence. - **Confidence:** Moderate-to-high. - **Scope and conditions:** Emergency manual override isn't always safe; some states require controlled escalation or an expert intermediary. - **Supports:** C-012, C-015. - **Links:** [[3. Resources/MOCs/AI Agent Safety MOC]]. ### Structure-note candidates - **Book map:** this note's Prompt 2 and Prompt 13 visuals. - **Theme map:** delegation → authority → observability → correction → transfer → evaluation. - **MOC adjacency:** [[3. Resources/MOCs/AI Agent Safety MOC]] for authority, verification, and control; [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] for skill, labor, and role change. - **Project routing:** No project link is justified without the reader's stated use. - **Unresolved-question map:** Q9–Q14 and Q22–Q24 are the most actionable for current agents. ### Attribution audit - The 36 source selections are `system-selected`; none is a reader highlight. - C-001–C-025 are paraphrased author claims with source anchors. - Confidence updates, grades, repairs, DEIB findings, current comparisons, and permanent notes are system analysis. - Noessel's 2024 agentive/agentic distinction is later author context, not part of the 2017 EPUB. - The Mermaid diagrams reconstruct and critique the argument; they are not book figures. **Do not attribute to Noessel:** the authority-contract model; “handoff debt”; supervision-cost value formula; current LLM-agent security findings; disability and design-justice critique; post-2017 regulatory analysis; Bayesian judgments; permanent-note wording; or any claim labeled **We infer** or **We conclude**. ## Prompt 11 — Teaching pack ### Audience and outcomes - **Audience:** Product designers, product managers, engineers, researchers, risk owners, accessibility practitioners, and technical communicators working on systems that act beyond a direct request. - **Prerequisites:** Basic product lifecycle and evaluation knowledge; no machine-learning math required. - **Learning objectives:** Learners will distinguish delegation modes, map lifecycle touchpoints, identify handoff risk, specify an independent evaluation, and critique whose interests “on behalf” represents. ### Vocabulary Agentive · agentic · assistive · automation · persistence · trigger · exception · constraint · directability · handoff · takeback · calibrated trust · measured value · external oracle. ### Lesson sequence 1. **Frame the category:** Compare a tool, assistant, deterministic automation, and agent. 2. **Map the lifecycle:** Setup → running → problems → handoff/takeback → disengagement. 3. **Inspect authority:** Goals, permissions, constraints, logging, revocation, and affected parties. 4. **Design failure:** False positives, false negatives, hostile inputs, resource limits, and graceful exit. 5. **Practice control transfer:** Context, readiness, acceptance, and recovery. 6. **Evaluate:** Results, calibrated confidence, cooperation, perceived value, measured value, security, rights, and distribution. 7. **Critique the future:** Separate 2017 forecasts from current evidence and law. ### Discussion and misconception checks 1. When does “on behalf of” become a testable claim? 2. Is a chatbot agentive if it acts only during a session? 3. Can a system be agentive to a purchaser and coercive to a worker? 4. What reveals a false negative? 5. When is human override unsafe? 6. Which metric could make a harmful agent look valuable? 7. Does a trustworthy interface prove a trustworthy system? 8. What must happen when the user dies, leaves, or revokes access? Misconceptions to test: - “Agentive” means conversational or human-like. **Correction:** The book centers persistent action for a goal. - More monitoring always improves safety. **Correction:** Passive vigilance decays and interruptions impose costs. - User confidence is reliability. **Correction:** Confidence needs calibration against observed results. - A test drive proves live safety. **Correction:** Simulated stakes and environments may not transfer. - Smarter automatically means safer. **Correction:** Accuracy, authority, reach, and attack surface move separately. - Human-in-the-loop means accountable. **Correction:** The human needs time, context, skill, authority, and a real choice. ### Activities and cases - **Lifecycle mapping:** Map an email triage agent against Appendix A's touchpoints. - **Failure table:** List false positives, false negatives, hostile inputs, and external harms for a calendar-booking agent. - **Handoff drill:** Design a staged takeover for a financial or site-management agent; identify context the human lacks. - **Baseline exercise:** Define measured value for an agent and include supervision cost. - **Power audit:** Identify purchaser, operator, data subject, bystander, worker, provider, regulator, and appeal owner. - **Repair exercise:** Rewrite C-022 so it becomes testable and risk-bounded. ### Reflection and assessment - Reflection: What work should remain deliberately human, and why? - Short assessment: Classify five systems by mode and defend the boundary. - Applied assessment: Produce a lifecycle, authority, failure, accessibility, and evaluation plan for one agent. - Rubric: source attribution, affected-party coverage, enforceable controls, baseline quality, uncertainty, and recovery design. ### Facilitator cautions and adaptations - Don't teach the category boundary, forecasts, or reviewer repairs as settled facts. - Disclose that the EPUB's figures lack descriptive alt text; provide a text lifecycle list before diagrams. - Avoid treating “human” as one role; distinguish operator, expert, affected person, and accountable owner. - Use cases outside cars, finance, homeownership, and U.S. consumer software. - Offer text, table, audio description, and keyboard-accessible diagram alternatives. - Explain acronyms and historical references; separate sentence difficulty from reference density. - The slavery/personhood discussion can divert attention from present human labor and racial history; frame it carefully or omit it when it doesn't serve the objective. ### Further reading - Bainbridge, “Ironies of Automation.” - Parasuraman, Sheridan, and Wickens, “A Model for Types and Levels of Human Interaction with Automation.” - Amershi et al., “Guidelines for Human-AI Interaction.” - NIST AI Risk Management Framework 1.0 and Generative AI Profile. - Costanza-Chock, *Design Justice*; Noble, *Algorithms of Oppression*; Benjamin, *Race After Technology*; Hamraie, *Building Access*. ## Prompt 12 — Action, PARA, experiments, and drift monitoring ### Action inventory **Take** - Add lifecycle, handoff, disengagement, and independent-oracle questions to agent reviews. - Separate capability, reliability, user confidence, perceived value, and measured value. - Require a baseline and count supervision and recovery cost. - Treat permission as enforced authority with revocation and audit. **Consider** - Use the book's touchpoint catalog during scenario planning. - Use agentive and assistive as modes within tasks rather than fixed product labels. - Add affected-party and distributional outcomes to the author's evaluation taxonomy. **Avoid** - Don't infer safety from intelligence, polish, or user confidence. - Don't rely on passive human monitoring as a catch-all control. - Don't treat a simulated test drive as proof of live performance. - Don't publish the review's permanent-note candidates as the author's conclusions. > [!danger] Most likely misuse > Turning the lifecycle checklist into proof that an agent is safe. The checklist finds questions; only enforced controls, independent tests, outcome data, and rights-respecting governance can answer them. **Needs more evidence:** Cross-cultural transfer, disability-led redesign, labor outcomes, current product performance, and long-term competence effects. ### PARA routing | Bucket | Material | Rationale | |---|---|---| | Projects | None assigned | 🔶 No current project was supplied. | | Areas | Agent/product design review practice | Ongoing responsibility if the reader owns agent systems. | | Resources | Lifecycle patterns, claim table, teaching pack, visuals | Reusable reference material. | | Archives | This complete book review | The source review belongs in the requested Book reviews archive. | ### Reversible experiments | Experiment | Claim | Hypothesis | Action | Measure | Duration | Stop condition | Review date | |---|---|---|---|---|---|---|---| | Lifecycle gap check | C-006 | The map finds missed states in an existing design | Review one agent spec against five lifecycle phases | New material gaps and severity | One review cycle | No new actionable gaps | 2026-09-24 | | Baseline-first evaluation | C-014 | A baseline changes the value judgment | Define incumbent, agent, and supervision-cost measures before a pilot | Outcome delta and review disagreement | One pilot | Metrics can't be collected ethically or cheaply | 2026-10-01 | | False-negative sample | C-009, C-013 | Independent sampling finds misses the agent can't see | Audit a small random sample of unflagged cases | Miss rate and harm class | Four weeks | Sensitive data or review burden exceeds approval | 2026-10-08 | | Handoff tabletop | C-011–C-012 | A rehearsal exposes missing context and authority | Run a disclosed tabletop with operators | Time to understand, accept, act, recover | One session | Exercise creates operational risk | 2026-09-17 | | Capability/authority split | C-022 | Risk rises when reach grows even if accuracy improves | Compare same model read-only vs bounded write access in a sandbox | New failure paths, reversible errors, blocked actions | Two test rounds | Any escape from sandbox | 2026-09-24 | These are design evaluations, not covert studies. Participants should know what is being tested. ### Decision rules - **Adopt** a pattern when it finds a material risk or improves measured recovery without shifting harm. - **Adapt** when it works only for a risk tier, user group, or task state. - **Test further** when confidence improves but affected-party outcomes remain unknown. - **Reject** when the pattern increases interruption, false confidence, inequity, or unsafe override. - **Defer** when no independent oracle or ethical data access exists. - If rejected, revise the permanent-note candidate tied to that experiment rather than preserving it as timeless advice. ### Drift monitoring | Claim or context | What can change | Sources | Frequency | Reassess by | Early trigger | |---|---|---|---|---|---| | C-017 agent infrastructure | Tool protocols, multi-agent systems, identity standards | NIST, standards bodies, primary research | 6 months | 2027-02-27 | Major agent protocol or deployment shift | | C-018 equity | Access, pricing, law, outcome evidence | Regulators, civil-rights research, service data | Annual | 2027-08-27 | New high-impact deployment or enforcement action | | C-020 labor | Adoption and task displacement | BLS, OECD, peer-reviewed labor studies | Annual | 2027-08-27 | Sector-specific displacement evidence | | C-022 safety | Agent capability, attacks, controls, evaluation | NIST, AI safety reports, primary security research | 6 months | 2027-02-27 | New attack class or reliable mitigation | | Legal context | EU AI Act implementation and other applicable rules | EUR-Lex and regulators | 6 months | 2027-02-27 | New binding guidance or deadline | Low-drift claims that don't need routine re-review: passive vigilance can decay; repeated task disuse can weaken skill; confidence and correctness differ; an independent baseline improves evaluation; Noessel's 2017 claims are fixed. ## Prompt 13 — Visual thinking pack ### Visual coverage matrix | Requested form | Disposition | |---|---| | Book map | Prompt 2's source-anchored part and chapter tables; a second diagram would repeat them. | | Argument map | Visual 1 below. | | Claim dependency graph | Mermaid in Prompt 4. | | Evidence map | Visual 4 below. | | Terminology map | Prompt 3's controlled-vocabulary matrix; a graph would imply a hierarchy the source doesn't establish. | | Causal-loop diagram | Omitted: social feedback claims lack enough evidence for a defensible loop. | | Timeline | Prompt 8's intellectual-lineage Mermaid covers the useful chronology. | | Stakeholder map | Omitted until a specific deployed agent supplies real actors and authority. | | Systems map | Visual 2 serves as the lifecycle system and control loop. | | Comparison matrix | Prompt 3's controlled vocabulary and Prompt 7's claim evaluation cover the comparison. | | Decision tree | Visual 3 below. | | Concept map | Visual 1 also serves as the concept map because the main concepts follow the argument spine. | | Action map | Prompt 12's decision rules and experiments are clearer as tables. | ### Visual 1: Book argument map - **Purpose:** Show how the definition supports the lifecycle and where broad future claims branch away. - **Nodes and groups:** Definition, value premise, lifecycle, evaluation, social effects, professional practice. - **Connections:** Solid arrows for direct argument; dashed arrows for weak or speculative transitions. - **Uncertainty:** C-002, C-021, and C-022 use dashed styling. - **Sources:** C-001–C-025; S-002–S-036. - **Layout:** Left-to-right argument flow with practical and speculative branches. ```mermaid flowchart LR DEF["Persistent delegated work\nC-001"] --> ROLE["Task-doer to task-manager\nC-003, C-006"] ROLE --> LIFE["Lifecycle design\nC-007–C-015"] LIFE --> EVAL["Independent evaluation\nC-013–C-014"] EVAL --> PRACTICE["Shared practice and metrics\nC-024"] ROLE --> SCALE["Wide adoption\nC-016–C-020"] SCALE --> PRACTICE LADDER["Tool-to-agent trajectory\nC-002"]:::uncertain -.-> DEF SMART["Smarter means safer\nC-022"]:::weak -.-> SCALE AGI["Rules seed AGI ethics\nC-021"]:::weak -.-> PRACTICE classDef uncertain stroke-dasharray: 5 5 classDef weak stroke:#b33,color:#b33,stroke-dasharray: 5 5 ``` ### Visual 2: Agentive lifecycle and control loop - **Purpose:** Reconstruct Appendix A in accessible text and show authority at every phase. - **Nodes:** Setup, running, problems, handoff, takeback, disengagement, external audit. - **Connections:** Normal flow, exception escalation, recovery, and exit. - **Grouping:** Human authority above; agent work below; audit outside both. - **Uncertainty:** Timing and responsible role are domain-specific. - **Sources:** C-007–C-015; S-014–S-031. - **Layout:** Circular state flow with audit as an external cross-check. ```mermaid stateDiagram-v2 [*] --> Setup Setup: capabilities, limits, goals, permissions, test Setup --> Running: authorized launch Running: monitor, notify, play, measure Running --> Problems: trigger, behavior, resource, or hostile-input failure Problems --> Running: tune and verify Problems --> Handoff: agent cannot safely continue Handoff --> HumanControl: context, readiness, acceptance HumanControl --> Takeback: stability restored Takeback --> Running: confirmed return Running --> Disengagement: goal ends or consent revoked HumanControl --> Disengagement Disengagement --> [*] ``` **External audit:** Samples outcomes across Running, Problems, Handoff, and Takeback. It must use evidence outside the agent's own model. ### Visual 3: Risk-tiered delegation decision tree - **Purpose:** Decide whether to use a tool, assistance, bounded agent, or no delegation. - **Nodes:** Repeatability, harm, oracle, reversibility, handoff readiness, enforced authority. - **Connections:** Each “no” narrows autonomy or stops deployment. - **Uncertainty:** Thresholds need domain owners and affected-party input. - **Sources:** C-004, C-007, C-011–C-015, C-022; S-014, S-021, S-025, S-029. - **Layout:** Top-down decision tree. ```mermaid flowchart TD A{"Recurring goal with observable conditions?"} -->|No| TOOL["Use a tool or ordinary workflow"] A -->|Yes| B{"Can success and harmful failure be independently observed?"} B -->|No| ASSIST["Keep human primary; use assistance only"] B -->|Yes| C{"Are actions reversible and low-impact?"} C -->|Yes| BOUND["Bounded agent with logs, undo, and sampling"] C -->|No| D{"Enforced scope, expert handoff, and tested recovery?"} D -->|No| STOP["Don't delegate action"] D -->|Yes| PILOT["Risk-tiered pilot with independent oversight"] ``` ### Visual 4: Excalidraw evidence map specification - **Purpose:** Separate author claims, book evidence, independent evidence, and system conclusions. - **Canvas:** Four vertical lanes: Claims → Book support → Independent checks → Current verdict. - **Nodes:** C-001–C-025 cards; E-001–E-012 evidence cards; verdict cards from Prompt 7. - **Connections:** Solid green for support, red for contradiction, gray dashed for indirect relevance. - **Grouping:** Practical lifecycle claims centered; speculative claims in a lower red-outlined zone. - **Labels:** Grade, independence, source year, and scope. - **Uncertainties:** Missing reader emphasis and missing representative outcome data appear as amber banners across the canvas. - **Sources:** All claim, evidence, and source-selection IDs. - **Recommended layout:** 16:9 landscape, swimlanes, legend in upper right, C-022 and C-021 visibly isolated. ### Omitted visuals - **Timeline:** The intellectual-lineage map in Prompt 8 already captures the useful chronology. - **Stakeholder map:** Better built for a specific agent and deployment; the book doesn't supply one concrete live system. - **Causal-loop diagram:** Broad social feedback claims lack enough evidence for a defensible causal model. - **Comparison matrix:** The controlled vocabulary and claim evaluation already provide the comparison at lower reading cost. - **Action map:** Prompt 12's decision rules and experiments are clearer as tables. ## Final synthesis ### Bottom line Noessel's durable contribution is a lifecycle vocabulary for technology that keeps working when the person isn't attending to it. The best chapters ask product teams to design the moments conventional happy-path UX overlooks: capability limits, permission, quiet operation, false negatives, correction, degraded performance, transfer of control, disengagement, and proof of value. Later HAI guidance and risk frameworks independently converge on much of that core. The category argument is less durable. “Agentive” works as a design lens, but the boundary with automation and assistance is porous, and current “agentic AI” describes backend planning and tool use that can span domains. The book's product examples demonstrate possibility, not comparative benefit. Its broadest safety claim fails: a more capable system can make fewer ordinary mistakes while reaching more systems, taking higher-impact actions, or accepting new adversarial inputs. Use the book to ask better design questions. Pair it with enforceable authority, threat modeling, independent evaluation, accessibility, affected-party governance, and current law before allowing an agent to act. ### Corrections and tensions retained - The EPUB gives no edition statement; “first edition” wasn't inferred from publication date. - The extractor initially missed 19 endnote bodies stored in `div.footnote`; it was corrected and the full source rerun. - Chapter 11 endnote 1 is truncated in the EPUB and remains marked as a source defect. - Prompt 10's default separate-note requirement conflicts with the user's one-file contract. Candidates stay embedded; no adjacent files or reciprocal MOC edits were made. - The canonical session-record file is also embedded here to preserve the one-file contract. ## Session completion check 1. **Prompts run:** 0–13. 2. **Usable outputs:** All prompts produced usable output. 3. **Adaptations:** Prompt 0 used the local EPUB after no exact Readwise match; Prompt 1 marks reader inputs unresolved; Prompt 10 embeds note candidates; Prompt 13 describes raster limits and omits unsupported visuals. 4. **Missing source material, ranked by consequence:** 1. Reader purpose, beliefs, and emphasis — high consequence for personalization, no effect on author-claim extraction. 2. Representative outcome studies for many product and social claims — high consequence for causal confidence. 3. Full transcription of 79 non-core raster images — moderate for accessibility and examples; the two core Appendix A diagrams were inspected. 4. Complete bibliography and the truncated Chapter 11 endnote — moderate-to-low for traceability. 5. Current status of every 2017 product example — low for the pattern analysis, high if cited as present fact. 5. **Assumptions:** Analytical depth; Archive PARA route; inferred current-agent design purpose; source-selected passages; selective-deep-read recommendation. 6. **Next prompt:** None; chain complete. 7. **YAML:** Required workflow fields updated and internally consistent. 8. **Attribution:** Author, third-party evidence, current context, and system analysis remain separated. 9. **Unrequested work not done:** No Reader publication, highlight creation, tag change, separate permanent note, MOC edit, or adjacent vault artifact. ## Work the Book — prompts used this session ```yaml title: "Work the Book — Prompts Used This Session" purpose: "Record the exact prompts executed during this Work the Book session so they can be reviewed, revised, or rerun." book_title: "Designing Agentive Technology" book_author: "Christopher Noessel" book_year: "2017" source_of_truth: "/Users/courtneyrobertson/Documents/2nd Brain/6. Templates/Books/Work the Book.md" source_spec_version: 2 working_note: "/Users/courtneyrobertson/Documents/2nd Brain/4. Archives/Book reviews/Designing Agentive Technology.md" output_files: - "/Users/courtneyrobertson/Documents/2nd Brain/4. Archives/Book reviews/Designing Agentive Technology.md" executed_prompts: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13] skipped_prompts: [] remaining_prompts: [] session_date: "2026-08-27" session_status: "complete" ``` ## Sources ### Layer 1 - Supplied EPUB: *Designing Agentive Technology: AI That Works for People*, Christopher Noessel, 2017. SHA-256 `5de4beaedcd8bac30740a93e02b5a50ceea537839a7a0abee56e9c0d63f8f606`. ### Current and independent context - [Rosenfeld Media book page](https://rosenfeldmedia.com/books/designing-agentive-technology/) — publisher metadata, May 2017 publication, stated audience and purpose. - [Christopher Noessel, “Agentive Tech and Agentic AI: Cousins not adversaries” (2024)](https://christophernoessel.medium.com/agentive-tech-and-agentic-ai-cousins-not-adversaries-2a2fa37c3134) — later author distinction between user-facing agentive strategy and agentic backend orchestration. - [Amershi et al., “Guidelines for Human-AI Interaction” (CHI 2019)](https://www.microsoft.com/en-us/research/wp-content/uploads/2019/01/Guidelines-for-Human-AI-Interaction-camera-ready.pdf) — 18 guidelines validated through practitioner evaluation. - [NIST AI Risk Management Framework 1.0](https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10) — lifecycle risk, roles, human-AI configurations, and accountable governance. - [NIST, “Strengthening AI Agent Hijacking Evaluations” (2025)](https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations) — indirect prompt injection and unintended agent action. - [NIST, “Lessons Learned: Tool Use in Agent Systems” (2025)](https://www.nist.gov/news-events/news/2025/08/lessons-learned-consortium-tool-use-agent-systems) — tool-enabled action, security, and reliability risk. - [Regulation (EU) 2024/1689, the EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32024R1689) — risk-based obligations and human oversight for covered systems. - [International AI Safety Report 2026](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026) — current agent capability, reliability, autonomy, and evaluation limits. - [Parasuraman, Sheridan, and Wickens, “A Model for Types and Levels of Human Interaction with Automation” (2000)](https://pubmed.ncbi.nlm.nih.gov/11760769/) — function allocation across automation types and levels. - [Bainbridge, “Ironies of Automation” (1983)](https://www.sciencedirect.com/science/article/pii/0005109883900468) — out-of-loop and abnormal-condition problems in automated systems.