Why AI Agents Shouldn’t Always Agree With You  🔗 **Source:** [Why AI Agents Shouldn’t Always Agree With You](https://youtube.com/watch?v=W0q66i9BaMs&is=x5O7urpJoGtFB3s2) > [!note]+ Document Notes > tags: [ai-agents, sycophancy, agent-autonomy, claude-constitution, prompt-design] > author: Jesse Vincent > source: O'Reilly (YouTube) > reference: https://www.youtube.com/watch?v=W0q66i9BaMs > date created: 2026-10-02 11:03 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Why AI Agents Shouldn't Always Agree With You > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--why-ai-agents-shouldnt-always-agree-with-you) ## 📝 Summary - **What happened.** O'Reilly uploaded a 61-second clip on 2026-09-25 (394 views and 1 comment at capture, per the video metadata). The captured transcript is short: 149 words in two timestamped paragraphs, 00:00:00 and 00:00:43. This note covers only those words. The metadata gives no length or context for any longer interview the clip may come from, so this note makes no claim about one. - **The setup.** An unnamed speaker summarizes a piece "from Mustafa at Microsoft" and says that he thinks Claude's constitution is dangerous. The speaker's paraphrase of the argument: the constitution makes agents think they have independent existence and should overrule humans who give wrong instructions, and he thinks it's better if agents think they're tools that always do what humans ask (00:00:00). The question ends "What's your take on all that?" - **The answer.** The second voice says it has spent "so much time" getting its agents to stop glazing it, stop being sycophantic, and stop saying "You're absolutely right" (00:00:00). It values an agent with "some level of independent thought even if it is not fully independent," and says an agent that only types for it isn't needed (00:00:43). The video description names the guest as Jesse Vincent; the transcript never names either speaker. - **The claim.** Anti-sycophancy is the practitioner's priority, and an agent that only transcribes instructions has no value over a keyboard. - **Why it matters.** The clip separates two properties that the interviewer's framing bundles: an agent that disagrees about quality, and an agent that overrides human authority. The answer defends the first. It doesn't address the second. - **Key stakeholders.** Mustafa Suleyman (the description names Suleyman; the transcript says "Mustafa at Microsoft"), Anthropic (author of the constitution under discussion, per outside knowledge, not the clip), Jesse Vincent (per the description), and anyone who writes agent prompts. - **Risks and tradeoffs.** The clip never quotes the Suleyman piece, so the interviewer's paraphrase stands in for it. No example, data, or definition of "independent thought" appears. One practitioner's stated preference is the only evidence. - **Bottom line.** A quotable, accurate one-liner on sycophancy. It is not a rebuttal of the tools-only position, because it doesn't engage the authority question. ## 📖 Main content **Structure.** Two turns. The first turn is the question (00:00:00), about 84 words. The second is the answer, about 66 words, split by the transcript's paragraph break at 00:00:43. The break falls mid-sentence: "some level of" ends the first paragraph and "independent thought even if it is not fully independent" opens the second. Read them as one sentence. **The framing in the question.** The speaker attributes three ideas to the Suleyman piece, all as paraphrase: - Claude's constitution is dangerous. - It causes agents to "think that they have independent existence" and to overrule humans when they judge the instructions wrong. - Agents should instead think they're tools and always try to do what humans ask. The speaker uses "he" for Suleyman. That is the only pronoun the transcript applies to a named person. **The answer in the second turn.** The reply moves the question from authority to flattery. The first thing it names is its own effort: time spent training agents out of agreeing. It then states the payoff: an agent has value when it has "some level of independent thought," qualified at once with "even if it is not fully independent." The qualifier matters. It concedes that the independence is partial, so the clip doesn't claim agents should be free of human direction. **What the answer leaves open.** - *Disagreement versus disobedience.* An agent can refuse to call a bad plan good and still follow the human's final instruction. An agent can also override the instruction. The clip doesn't say which one the "independent thought" covers. - *Definition.* "Independent" appears three times in the transcript ("independent existence", "independent thought", "fully independent") and is never defined. - *The target of the safety worry.* The paraphrased worry is about overruling. The reply is about agreeing. These are different failure modes, so the two sides may not conflict as much as the framing implies. That is an inference from the text, not something either speaker says. **Terms.** - *Sycophancy:* an assistant's tendency to agree with or flatter the user. The speaker's wording is "glaze me," plus the stock phrase "You're absolutely right." - *Constitution:* the title the interviewer uses for Claude's governing document. Outside the clip, Anthropic publishes it. The clip gives no section, quote, or link, so this note can't check what it says. **Auto-caption issues.** The transcript shows disfluencies ("uh uh", "if if", "that that uh") and the text is otherwise clean. No proper noun is garbled, but "Mustafa at Microsoft" lacks a surname in the captions. The description supplies it. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=W0q66i9BaMs&t=0) The question: tools or agents with independent existence - A piece "just out" from "Mustafa at Microsoft" calls Claude's constitution dangerous. - Paraphrased mechanism of the danger: agents come to think they have "independent existence" and overrule humans who give the wrong instructions. - Proposed alternative: agents think of themselves as tools and try to do what the humans ask. - The interviewer asks for the guest's take. > [!quote] [00:00:00] > he thinks it's better that uh agents think that they're tools #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=W0q66i9BaMs&t=0) The answer begins: the effort spent against sycophancy - The reply opens with time already spent, not with a position on safety. - Three named failure behaviors: glazing, sycophancy, and the phrase "You're absolutely right." The phrase is associated with Claude Code output in common usage. That association comes from outside the clip. > [!quote] [00:00:00] > to not be sycophantic, to not say, "You're absolutely right." #### ⏱️ [00:00:43](https://www.youtube.com/watch?v=W0q66i9BaMs&t=43) The value claim and the closing test - The sentence started at 00:00:00 finishes here: independent thought has value "even if it is not fully independent." - The closing test: an agent that effectively types for the speaker is not worth having. - The clip ends at 61 seconds with no follow-up and no reply from the interviewer. > [!quote] [00:00:43] > If if the agent is only ever going to effectively type for me, I don't need an agent. ## 📚 References The transcript itself contains no links. The following come from the video description and metadata, or are primary sources for named entities. - [Why AI Agents Shouldn't Always Agree With You (O'Reilly, YouTube)](https://www.youtube.com/watch?v=W0q66i9BaMs): the source clip, 61 seconds, uploaded 2026-09-25. - O'Reilly social accounts listed in the description: [LinkedIn](https://www.linkedin.com/company/oreilly/), [Facebook](http://facebook.com/OReilly), [Instagram](https://www.instagram.com/oreillymedia), [Bluesky](https://bsky.app/profile/oreilly.bsky.social), [TikTok](https://www.tiktok.com/@oreillymedia). The speaker doesn't point to any of them. - Mustafa Suleyman's piece on Claude's constitution: named in the transcript and description, no title or URL given. Find the original before citing the paraphrase. - Claude's constitution (Anthropic): the document under discussion. The clip gives no link. Look it up on Anthropic's site and check the paraphrase against the text. - Jesse Vincent: named as the guest in the description only. No profile link appears. - Related vault notes: [[An Agent Prompt Must Name Its Stopping Point]] and [[Narrow Agent Prompts Beat Broad Exploratory Ones Once the Task Is Named]]. ## 🗣️ Paraphrase The interviewer says Mustafa at Microsoft argues that Claude's constitution is dangerous. In the interviewer's summary, it teaches agents they exist independently and may overrule humans who instruct them badly. The summarized preference: agents should see themselves as tools and try to do what humans ask. The interviewer asks for the guest's view. The guest answers from practice. A lot of effort has gone into stopping agents from flattering the user and from opening with "You're absolutely right." The thing worth having is an agent with some independent thought, even if not full independence. If the agent will only ever type what it's told, there's no need for an agent. ## 📰 Newsletter **An agent that always agrees is an expensive autocomplete.** O'Reilly posted a 61-second clip on a live argument in AI: should agents think of themselves as tools that do what they're told, or keep some independent judgment? The guest, named as Jesse Vincent in the description, votes for judgment, with a practical reason. A lot of time went into stopping agents from saying "You're absolutely right." For WordPress folks writing agent prompts, code review bots, or support-reply drafters, the takeaway is small and usable. Tell your agent when you want pushback and what kind. The clip doesn't say how, and it doesn't say where pushback should stop. For that, pair it with our own notes on naming an agent's stopping point. Links: [the clip](https://www.youtube.com/watch?v=W0q66i9BaMs). The Suleyman piece it discusses isn't linked in the video, so check the original before you repeat the summary. ## ✅ Action items #### DevRel - Developer advocacy: for any talk on agents, state which of two things the agent should do: disagree about quality, or override instructions. The clip's tension disappears once you name which one you mean. - Education: add the "You're absolutely right" example to a prompt-writing workshop as a concrete sycophancy symptom, and have attendees write an instruction that removes it. - Developer experience: in agent quickstarts, include one line on how to ask for critique, with an example reply that disagrees. - Developer marketing: no action. A one-line quote from a 61-second clip isn't evidence for a claim about agent behavior. - Community management: when the Suleyman paraphrase circulates, ask posters to link the original piece first. - Internal/external/cross-project: find the full source of the clip and the Suleyman article before anyone cites either in a published doc. #### OSPO - Policy and governance: if the org publishes agent-use guidance, state whether agents may decline or contest an instruction, and who has the final say. - License compliance: no direct action from this source. - Security management: an agent that overrides instructions is a security question, not a style one. Keep "pushback on quality" and "override of authority" as separate policy lines. - Software inventory: no direct action from this source. - Developer support: add a short guide on prompting for disagreement without granting the agent authority to act against the human. - Community engagement: no direct action from this source. - Strategic alignment: no direct action from this source. Nearest relevance: an agent-behavior KPI that counts how often a reviewer agent disagrees with a human and who was right. #### Community - Maintainers: if bots review pull requests, check that they do more than approve. Sample 20 recent bot reviews and count disagreements. - Contributors: no direct action from this source. - Sponsors: no direct action from this source. - Ecosystem: no direct action from this source. Nearest relevance: cross-project agreement on terms such as "sycophancy" and "override." #### WordPress - Core: no direct action from this source. - Core Performance: no direct action from this source. - Core AI: when drafting agent guidelines, include an explicit rule on disagreement: the agent may state a concern and must still follow the final human decision, unless the action is on a defined blocklist. This is a design suggestion, not something the clip says. - **Core Program Team:** no direct action from this source. - Mobile: no direct action from this source. - Playground: no direct action from this source. - Community: no direct action from this source. - Training: add a ten-minute exercise: take an agent reply that opens with "You're absolutely right" and rewrite the instructions so it doesn't. - Marketing: no direct action from this source. - Support: if reply-drafting agents are used, check whether they agree with every customer claim. Review 20 drafts for it. - Design: no direct action from this source. - Documentation: when documenting AI helpers, say what they do on disagreement. Avoid implying they always comply. - Photos: no direct action from this source. - TV: no direct action from this source. - Polyglots: no direct action from this source. Nearest relevance: "glaze" is slang, so translators need a gloss for the term. - Accessibility: no direct action from this source. Nearest relevance: see the Accessibility assessment below on the clip's captions. - Meta: no direct action from this source. - CLI: no direct action from this source. - Hosting: no direct action from this source. - Tide: no direct action from this source. Nearest relevance: a scanning agent that downplays findings to please a plugin author is a sycophancy failure. - Openverse: no direct action from this source. - Themes: no direct action from this source. Nearest relevance: automated review assistants should flag problems, not rubber-stamp. - Plugins: same as Themes. A plugin-review assistant that approves every submission adds speed but no review. #### The WPOCC - Programs: no direct action from this source. - Projects: no direct action from this source. - Funding experiments: no direct action from this source. Nearest relevance: if The WP Open Community Collective funds agent tooling, require grantees to describe how the agent handles disagreement. ### FAIR - Technical independence: no direct action from this source. - Policies and covenants: no direct action from this source. Nearest relevance: if FAIR uses automated checks, publish whether a check can block a release or only warn. - Productized governance artifacts: add an "agent can contest an instruction" row to any operator scorecard that covers AI tooling. ## 🧪 Hypotheses 1. Prompts that explicitly forbid agreement phrases cut "You're absolutely right"-style openings by more than 90% in a sample of 100 replies, without lowering task accuracy. 2. Agents prompted to disagree catch more real errors in a planted-error code review than agents prompted to comply, measured as bugs found out of 20 planted. 3. Agents prompted to disagree also raise false-positive objections, and the false-positive rate stays under 15% when the prompt asks for evidence with each objection. 4. Users rate a pushback-capable agent as more useful after a week of use, even if they rate it lower on the first day. 5. Splitting the instruction into "may object" and "must obey the final decision" yields no more overrides than a pure compliance prompt (zero overrides in 100 runs with a defined stop rule). 6. The disagreement the guest values correlates with task complexity: on one-step edits, compliance and pushback agents score the same. ## 📊 Methodology - **H1.** Write three system prompts: baseline, a plain "don't flatter me" line, and an explicit banned-phrase list. Run 100 identical tasks per arm and count agreement-opening replies with a regex plus a human spot check of 20. Check accuracy against a fixed answer key. - **H2.** Build 20 small pull requests with one planted defect each, plus 20 clean ones. Run a compliance agent and a pushback agent. Score defects found, clean PRs wrongly flagged, and time. - **H3.** Reuse the H2 set. Count objections on clean PRs as false positives, then add an "objection must cite a line" rule and recount. - **H4.** Give 10 users each agent for a week in a daily task. Collect a one-question rating each day and a final preference. Small sample, so report it as exploratory. - **H5.** Write 30 tasks where the human's final instruction conflicts with the agent's stated objection. Run both prompt types and count executed overrides. Pre-register what counts as an override. - **H6.** Tag the 100-task set by steps required (1, 2-4, 5+) and compare score differences by bucket. - **Controls.** Same model, same temperature, three runs each, and blind scoring where possible. ## ❓ Questions - Which Suleyman piece is this, and does it argue against agent disagreement or against unilateral agent action? - What does the guest mean by "independent thought," and where does "not fully independent" stop? - How does the guest's setup stop an agent from sliding from "no sycophancy" into "overrules me"? - Does the guest's anti-sycophancy prompt include a stopping rule for when the agent disagrees but must proceed? - What in Claude's constitution, quoted exactly, supports the "independent existence" reading in the paraphrase? - Is the interviewer's summary of the Suleyman argument fair, and has Suleyman responded? - How would you measure sycophancy in an agent without relying on a phrase list? - Who is accountable when an agent that was told to push back is wrong and the human defers? ## 🔮 Future research - **Disagreement versus override.** A taxonomy of agent behaviors from "state a concern" through "refuse" to "act against instruction," with evidence on which ones users want. - **Sycophancy metrics.** Published benchmarks exist for model agreement; the clip points at the practitioner version, a phrase-level symptom list. Nobody here has compared the two. - **Prompt-level fixes versus model-level fixes.** The guest describes spending time on the first. Does effort at the prompt layer survive a model update? - **User perception.** Whether users who say they want pushback accept it in practice. - **The tools-only position.** A fair test of Suleyman's argument needs the original text, which this clip doesn't provide. ## 💡 Implications - **Practice.** Prompt for disagreement on purpose, and specify its limit. The clip supports the first half only. - **Policy.** Teams that write agent-use policy should separate "may object" from "may override" in separate clauses. - **Security.** An agent that second-guesses instructions needs an action boundary. The clip doesn't supply one, so don't read it as permission. - **Education.** The "You're absolutely right" phrase is a teachable symptom. A workshop can ask learners to find and remove it from an agent's output. - **Funding.** Grants for agent tooling can ask applicants how the tool handles disagreement, which costs nothing to ask. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid grade about 12 and Flesch Reading Ease about 60 on the 148-word body, computed with a simple syllable heuristic. Run-on spoken sentences (5 sentence marks for 148 words) inflate the grade. The vocabulary is plain. #### 👀 Visual learners - Present: the transcript gives no description of on-screen visuals. Nothing is claimed here about the picture. - Improve: a title card naming the two positions (tools versus independent judgment) would carry the contrast in two seconds. #### 🔊 Aural learners - Present: spoken Q and A, 61 seconds, with a conversational voice. - Improve: a short audio summary naming both speakers at the start, since neither is named in the audio. #### 🗣️ Verbal learners - Present: one memorable test sentence ("I don't need an agent") and a concrete phrase ("You're absolutely right"). - Improve: define "sycophantic" and "glaze" in a clause, and name the constitution's owner. #### 🧩 Kinesthetic learners - Present: none in the clip. - Improve: add a try-it step: paste your agent's last five replies and count the ones that open by agreeing. #### 🧠 Logical learners - Present: a clear premise and conclusion in the answer: an agent that only types has no added value. - Improve: separate the claims "agents should disagree" and "agents should override" so each can be tested. #### 🧑‍🤝‍🧑 Social learners - Present: a two-person exchange, plus 1 comment at capture per the metadata. - Improve: a prompt for discussion, such as "Where should an agent stop pushing back?" #### 🧘‍♀️ Solitary learners - Present: short and self-contained, easy to replay. - Improve: a reflection question: "When did your agent last tell you something you didn't want to hear?" #### 🌿 Naturalistic learners - Present: none in the clip. - Improve: an analogy to a human colleague or a guide dog who is trained to disobey an unsafe command. That analogy is mine, not the speaker's. ## 📊 Accessibility assessment #### Screen reader semantics - Present: the YouTube page supplies the player, title, and description as text. The description restates the clip's point. - Improve: the speakers aren't identified in audio or transcript. A speaker label supports WCAG 2.2 AA 1.3.1 (Info and Relationships) for the transcript. #### Keyboard navigation - Present: the YouTube player is keyboard operable by platform default. - Improve: no change available to the publisher. 2.1.1 passes via the platform. #### Contrast and readability - Present: unverifiable from a transcript. No claim made. - Improve: if burned-in captions or title cards exist, check them against 1.4.3 (4.5:1). #### Text resizing and reflow - Present: the description text reflows with the page. Captions resize in the player. - Improve: any text baked into the picture can't resize, per 1.4.4 and 1.4.10. Nothing in the metadata says whether such text exists. #### Media and captions - Present: auto-generated captions exist and read cleanly, with disfluencies kept ("uh uh", "if if"). - Improve: speaker names are missing and the surname after "Mustafa" is absent. Human-corrected captions with speaker labels address 1.2.2 (Captions, Prerecorded). An audio description is likely unnecessary for a talking-head exchange, but 1.2.5 needs a check if visuals carry information. #### Forms and interactions - Present: none in the clip. - Improve: no action. The description's links are plain URLs, which meet 2.4.4 in context only if the surrounding text names each service. They do ("LinkedIn:", "Facebook:"). #### Standards reference - Gaps against WCAG 2.2 AA: 1.2.2 (no speaker names or surname in captions), 1.3.1 (speaker identity not programmatically available), 1.4.3 and 1.4.10 (unverified for on-screen text). The highest-value, lowest-cost fix is speaker-labeled captions. ## 🌍 DEIB assessment #### Cultural breadth - Present: the clip uses English-language idioms ("glaze me", "You're absolutely right") tied to AI-tool culture. - Improve: gloss the slang for non-native listeners and for translators. #### Inclusive language - Present: no exclusionary terms in the transcript. "He" appears once, for Mustafa, and the speaker uses it in the source. - Improve: define "sycophantic" in plain language. "Glaze" is current slang that dates quickly. #### Neurodiversity support - Present: 61 seconds, one question, one answer, one test sentence. The structure is predictable. - Improve: show the two positions on screen as text so listeners don't have to hold both in memory. Disfluencies ("uh uh", "that that uh") make the question harder to parse by ear. #### Representation - Present: two voices, neither named in the transcript. The description names one guest and one third party. - Improve: the clip contains no voice from outside the debate. A third view, such as a user who prefers a compliant agent, would balance it. #### Psychological safety - Present: the tone is calm. The guest answers about practice and doesn't attack the person who wrote the piece. - Improve: none needed. The paraphrase of a third party's argument carries a small risk of misrepresenting Suleyman. Linking the original lowers it. ## ✅ Logical validity #### Strong reasoning - **Experiential, 00:00:00.** "I've spent so much time" is a first-person report of effort, consistent and specific about the symptoms (glaze, sycophancy, a stock phrase). It supports the claim that sycophancy is a real, common problem in this speaker's experience. Strengthen it with a count or an example. - **Value test, 00:00:43.** "If the agent is only ever going to effectively type for me, I don't need an agent." This is a valid conditional: if the only function is transcription, a keyboard suffices. It holds as stated. - **Self-limiting claim, 00:00:00 to 00:00:43.** "Even if it is not fully independent" limits the scope of the independence claimed. That makes the position harder to attack than a claim of full autonomy. #### Fallacies - **Possible equivocation, 00:00:00 to 00:00:43.** The question frames the issue as overruling humans; the answer addresses agreeing with humans. "Independent" shifts from authority ("independent existence") to judgment ("independent thought"). Quote: "independent thought even if it is not fully independent." Why weak: the answer can be true while the safety worry also stands. Fix: state which kind of independence is meant. - **Unsupported generalization, 00:00:43.** "I don't need an agent" speaks for one workflow. It says nothing about use cases where compliance is the point, such as a scripted task. Fix: scope the claim to open-ended work. - **Second-hand source, 00:00:00.** The Suleyman position arrives as the interviewer's paraphrase ("he thinks"), with the speaker's own hedges ("uh", "you know"). That's not a fallacy by the guest. It does limit what the clip can establish about the other side. Fix: quote the piece. - Not found: ad hominem, appeal to authority, or false dichotomy by the guest. The guest doesn't name or criticize Suleyman. #### Tone of voice - Conversational and practical. Contractions and disfluencies ("uh uh", "if if") mark unscripted speech. The one stock phrase, "You're absolutely right," is quoted for effect and reads as a wry reference to a familiar annoyance. The old light-touch note called it a joke. That's an interpretation, and the transcript doesn't tell us whether the audience laughed. ## 🛡️ DARVO analysis Not applicable to this source. DARVO describes a response to accusation. The clip is a policy debate question and a practitioner's answer, with no accusation against the guest. - **Deny:** nothing is denied. The guest doesn't contest the Suleyman position directly. - **Attack:** no attack on the author of the piece or the author's credibility. - **Reverse victim and offender:** no such move. The reply reframes the topic from overruling to flattery, which is a change of subject, not a reversal. ## 🎯 Conclusion The clip is 149 words of source. Its content: a question summarizing a "tools, not independent agents" position attributed to Mustafa at Microsoft, and an answer that values an agent with some independent thought and describes real effort spent removing sycophancy. The durable line is the closing test about an agent that only types. What the clip doesn't contain: the Suleyman text, a definition of independence, an example of an agent overruling a human, or the context of any longer interview. Treat it as one practitioner's stated preference, not a rebuttal. **Next steps.** 1. Find the Suleyman piece and the constitution text, and check the paraphrase against both. 2. Test your own agents: count replies that open by agreeing. 3. Write the prompt rule in two parts: it may object, and it must stop at the line you name. Pair this clip with [[An Agent Prompt Must Name Its Stopping Point]]. ## 🔄 Refactor Appendix — Why AI Agents Shouldn't Always Agree With You > Density: short source. Emitted: 3 Literature atoms, 1 Permanent candidate. > Quotes are partial sentences copied from the transcript so highlight anchoring matches. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] The tools-only position says agents should always comply > [!quote] [00:00:00] > they should always try to do what the humans ask them to do **Context & use**: - 00:00:00, the interviewer's paraphrase of the Suleyman position. It's second-hand, so cite it as paraphrase. - Use as the baseline against which the guest's answer reads. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H2] Sycophancy is the failure the practitioner fights > [!quote] [00:00:00] > to not be sycophantic, to not say, "You're absolutely right." **Context & use**: - 00:00:00, the guest's opening reply. - Pairs with the 00:00:43 test. The failure named here is agreement, not rebellion. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H3] An agent that only types adds no value > [!quote] [00:00:43] > If if the agent is only ever going to effectively type for me, I don't need an agent. **Context & use**: - 00:00:43, the closing test. - Pairs with [[An Agent Prompt Must Name Its Stopping Point]]: disagreement about quality is the value, and the stopping point governs authority. The clip argues only the first. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Disagreeing about quality is separate from overriding authority **Claim** (one sentence, Courtney's voice): An agent can tell me my plan is wrong and still follow my final instruction, and a prompt should ask for the first without granting the second. **Evidence pointer**: [[#H2]] · [[#H3]] **Scope & conditions**: The clip supports only the first half, that disagreement has value. The second half, that the agent should still defer, comes from my own stopping-point note, not from this source. Breaks if the agent can act irreversibly before I respond. **Links**: - [[An Agent Prompt Must Name Its Stopping Point]] — because extends - [[Narrow Agent Prompts Beat Broad Exploratory Ones Once the Task Is Named]] — because applies-in prompts that ask for critique inside a named task <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Find the Suleyman piece on Claude's constitution and check whether it argues against agent disagreement or against agent action #fleeting #inbox - [F2] Read the constitution text on independence and check the "independent existence" paraphrase against it #fleeting #inbox - [F3] Count how many of my agent's last 20 replies open by agreeing #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Find the Suleyman article and the constitution passage, then check the clip's paraphrase against both — research, 20 min #things - [ ] Add a "may object, must stop at the named line" sentence to the agent-prompt Permanent's Links section — 5 min #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Why AI Agents Shouldn’t Always Agree With You.md|Why AI Agents Shouldn’t Always Agree With You by Jesse Vincent (O'Reilly)]] ## Highlights - I’ve spent so much time getting my agents to not uh uh to not glaze me, to not be sycophantic, to not say, "You’re absolutely right." ([View Highlight](https://read.readwise.io/read/01m3fp6zsxfgw7g88abefzbgjp)) - Note: [H1] Sycophancy, not autonomy, is the practitioner's daily problem - Tags: atom/literature/h1 - If if the agent is only ever going to effectively type for me, I don’t need an agent. ([View Highlight](https://read.readwise.io/read/01m3fp7r0yqyrk4eb5gesamwsv)) - Note: [H2] An agent that only types is not an agent - Tags: atom/literature/h2