The Next Horizon in Agents
🔗 **Source:** [The Next Horizon in Agents](https://twitter.com/btaylor/status/2077767352431882355/?rw_tt_thread=True)
> [!note]+ Document Notes
> tags: [ai-agents, long-horizon-planning, outcome-based-pricing, ai-moats, customer-experience]
> author: Bret Taylor
> source: X (Sierra announcement thread)
> reference: https://twitter.com/btaylor/status/2077767352431882355/?rw_tt_thread=True
> date created: {{date:YYYY-MM-DD}} {{time:HH:mm}}
>
> # 🏷️ Title: The Next Horizon in Agents
>
<!-- 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)
### 🔗 References
- [Horizon product page](https://sierra.ai/product/horizon)
- [Introducing Sierra (2024 launch post)](https://sierra.ai/blog/introducing-sierra)
- [Satya Nadella on ceding value to AI models](https://x.com/satyanadella/article/2066182223213293753)
## 📝 Summary
- **What happened:** Bret Taylor announced Horizon, Sierra's platform for agents that pursue long-horizon goals (loan origination, healthcare prior authorization, referral scheduling) across days, weeks, or months of customer interactions — the company's biggest expansion since its 2024 launch.
- **Why it matters:** It reframes enterprise agents from single-conversation customer service toward multi-step revenue outcomes, and pairs that with outcome-based pricing — customers pay for delivered business outcomes, not tokens.
- **Key stakeholders:** Sierra (Taylor and Clay Bavor), Agent OS enterprise customers (Santander, Rocket Mortgage, Cigna, "almost half of the Fortune 50"), design partners like a US healthcare provider, and competing agent platforms.
- **Risks and tradeoffs:** The "data moat" claim is asserted, not demonstrated; outcome-based pricing raises attribution questions (who decides which outcomes the agent caused?); proactive multi-month outreach agents raise consent, privacy, and regulatory exposure — especially in healthcare and lending, the two verticals named.
- **Bottom line:** A significant product-category claim — agents as outcome-driving systems with a context engine and long-horizon planner — wrapped in confident marketing whose strongest verifiable signal is the pricing model shift.
## 📖 Main Content
- Taylor's core argument: an individual conversation is a small ingredient in complex business outcomes. Existing platforms lack two things — a **context engine** to stitch interactions together, and **long-horizon planning** so the agent can reason *between* engagements about the next action.
- Horizon extends Agent OS (Sierra's conversational platform) with both. Agents can pursue goals — close a sale, upgrade a plan, schedule a test drive — instead of following rigid standard operating procedures.
- The learning loop: each closed or rejected sale feeds decision-making for future conversations. Taylor frames accumulated interaction memory as "a durable, expanding moat."
- Positioning against two current executive anxieties: differentiation (Nadella's "ceding value to AI models") and tokenomics (controlling AI spend). Sierra's answer to the latter: outcome-based pricing, with Sierra bearing token-spend risk.
- Uncertainties: no benchmarks, metrics, or completion rates are given; the healthcare referral example is a design partner, not a published result; "moat" and "compounds" are asserted mechanisms with no supporting data in the thread.
### 🗒️ Notes
- Announcement structure: what Horizon is → why (front-door thesis) → design-partner example → tokenomics/moats argument → outcome pricing close.
- "Digital front door" thesis: Taylor and Bavor claim the AI agent will eclipse website and mobile app as the primary customer engagement surface.
- Scale claim: Agent OS powers phone and chat for hundreds of companies including almost half of the Fortune 50.
- The two named workload examples (loan origination, prior authorization) are both heavily regulated processes — a notable choice signaling enterprise-compliance confidence.
- Outcome pricing inverts the cloud-era metered model: vendor absorbs variable inference cost, customer buys results. This is the thread's most concrete, falsifiable commitment.
## 📚 References
- [Sierra Horizon](https://sierra.ai/product/horizon) — primary product page.
- [Introducing Sierra](https://sierra.ai/blog/introducing-sierra) — 2024 launch context.
- [Satya Nadella article on X](https://x.com/satyanadella/article/2066182223213293753) — the "ceding value" framing Taylor responds to.
## 🗣️ Paraphrase
- Sierra launches Horizon, a platform where agents chase long-running goals like loans and prior authorizations rather than answering single conversations.
- These agents contact customers proactively over weeks or months, working out the steps to close a sale or schedule an appointment.
- Every interaction feeds the agent's memory; Sierra argues this accumulated context becomes a defensible business moat that competitors can't copy.
- Agents so far delivered experiences; Horizon claims to deliver outcomes.
- Sierra's Agent OS already runs conversations for hundreds of companies, including nearly half the Fortune 50.
- A healthcare design partner uses Horizon to schedule specialist referrals — dozens of texts and calls across patient, specialist, and physician that existing platforms can't stitch together.
- Horizon adds a context engine and long-horizon planning to Agent OS so agents pursue goals instead of following rigid scripts.
- Business leaders keep asking Taylor about differentiation and token spend; Horizon is Sierra's answer to both.
- Your customer context stays proprietary as frontier models improve, so you benefit from new models without helping competitors.
- You pay for business outcomes delivered, not tokens — Sierra absorbs token-spend management.
## 📰 Newsletter
Sierra just moved the agent goalposts: Horizon agents don't answer tickets, they chase outcomes — nudging a customer for weeks until the loan closes or the specialist appointment sticks. Two things worth WordPress folks' attention: the **context engine** framing (stitching many touchpoints into one goal-directed thread) is exactly what membership, LMS, and WooCommerce follow-up flows try to fake with drip plugins today, and the **pay-for-outcomes, not tokens** pricing is a shot across the bow of every metered AI add-on in our ecosystem. If your product roadmap includes "AI assistant," the bar just moved from answering questions to finishing jobs. Read the [Horizon announcement](https://sierra.ai/product/horizon) and ponder what an outcome-priced WordPress agent would even promise.
## ✅ Action Items
#### DevRel
- Draft a talk comparing conversation-scoped vs. goal-scoped agent architectures; use Horizon's context-engine framing as the foil.
- Add long-horizon agent patterns (state persistence, between-session planning) to developer education backlog.
- Track how outcome-based pricing lands with developers versus token pricing in community sentiment.
#### OSPO
- Policy & Governance: assess what an outcome-priced vendor contract implies for data ownership clauses — the "your context remains yours" claim needs contract language, not marketing.
- License Compliance: not applicable to this source — no code or licenses discussed.
- Security Management: catalog risks of agents holding months of customer interaction memory (breach blast radius grows with the moat).
- Software Inventory: track which internal tools accumulate cross-interaction context that would need SBOM-style ownership records.
- Developer Support: brief teams on vendor lock-in dynamics of proprietary context engines.
- Community Engagement / Strategic Alignment: watch for open-source context-engine equivalents worth upstream investment.
#### Community
- Maintainers: discuss whether open agent frameworks need a shared long-horizon state model to stay competitive.
- Contributors: good-first-issue candidates around session persistence in open agent tooling.
- Sponsors: outcome-based pricing is a fundable open-source research question (attribution models).
- Ecosystem: coordinate a shared vocabulary — "context engine" and "long-horizon planning" are about to be marketing-diluted.
#### WordPress
- Core: no direct action — monitor.
- Core Performance: not applicable to this source — no performance surface.
- Core AI: evaluate the goal-pursuit vs. rule-following distinction for the AI roadmap's agent abilities work; guardrails differ fundamentally.
- **Core Program Team:** note the cross-team dependency pattern: long-horizon agents touch privacy, consent, and messaging simultaneously.
- Mobile: proactive agent outreach implies push-notification consent flows worth reviewing.
- Playground: a sandboxed "agent pursues a goal across sessions" demo would teach the concept well.
- Community: meetup discussion prompt — what would outcome-driven agents mean for client services?
- Training: add a lesson distinguishing chatbots, copilots, and goal-pursuing agents.
- Marketing: watch "outcomes not tokens" phrasing; it will shape buyer expectations for all AI products.
- Support: proactive agents generate inbound confusion ("why is this bot texting me?") — anticipate deflection docs.
- Design: multi-session agent UX (where did we leave off?) is an unsolved design pattern worth research.
- Documentation: not applicable to this source — no APIs to document.
- Photos / TV / Polyglots / Openverse: not applicable to this source — no media, captioning, or translation surface.
- Accessibility: proactive multi-channel outreach (SMS + voice + chat) must remain accessible across all channels; flag for pattern review.
- Meta: not applicable to this source.
- CLI: not applicable to this source.
- Hosting: agents calling home over months = long-lived webhooks/queues; hosts should note the workload shape.
- Tide / Themes / Plugins: plugin review implications if outcome-priced AI plugins emerge (what claims are verifiable?).
#### The WPCC
- Fund a small research note on outcome attribution in agent pricing — directly relevant to sustainable open-source AI services.
#### FAIR
- Technical independence: the proprietary context-moat argument is the exact centralization pattern FAIR exists to counter; document it as a case study.
- Policies and covenants: outcome pricing without transparent attribution is an anti-capture review candidate.
## 🧪 Hypotheses
- H1: Agents with persistent cross-interaction context will close multi-step outcomes (e.g., scheduled referrals) at a measurably higher rate than stateless conversational agents.
- H2: Outcome-based pricing will shift vendor behavior toward conservative, high-confidence agent actions (risk transfer changes action selection).
- H3: Accumulated interaction memory produces diminishing returns as a moat once frontier models can few-shot from small context samples.
- H4: Proactive long-horizon outreach agents will trigger regulatory scrutiny (TCPA, HIPAA, fair-lending) faster in the named verticals than in retail.
- H5: Enterprises will over-attribute outcomes to agents in the first year of outcome-priced contracts, inflating reported ROI.
### 📊 Methodology
- H1: randomized rollout at a design partner — referral completion rate, stateful vs. stateless agent arms.
- H2: audit logs comparing action-selection distributions under token-priced vs. outcome-priced deployments.
- H3: ablation study — agent performance with full interaction history vs. sampled subsets across model generations.
- H4: longitudinal tracking of enforcement actions and consent-related complaints in healthcare/lending agent deployments.
- H5: independent attribution audit comparing vendor-reported outcomes to holdout-group baselines.
### ❓ Questions
- How is an "outcome" contractually defined and who adjudicates disputed attribution?
- Is customer interaction data portable if a client leaves Horizon — or is the "moat" really Sierra's?
- How does the long-horizon planner handle a customer who says stop — does goal pursuit yield to consent immediately across all channels?
- What failure rate is acceptable for an agent autonomously working a healthcare referral for weeks?
### 🔮 Future Research
- Attribution science for agent-driven outcomes (the pricing model depends on it, and the thread is silent on it).
- Consent frameworks for proactive, multi-month, multi-channel agent outreach.
- Whether context moats survive model-capability jumps — the moat thesis's key vulnerability.
### 💡 Implications
- **Practice:** "agent" is being redefined from conversational responder to goal-pursuing system; product roadmaps citing agents inherit the higher bar.
- **Policy:** outcome-priced AI in lending and healthcare will collide with existing sector rules on solicitation and disclosure before general AI regulation catches up.
- **Security:** months of accumulated customer context concentrates breach impact; the moat is also the honeypot.
- **Education:** curricula still teach single-turn prompting; long-horizon planning and state management are the emerging skills gap.
- **Funding:** outcome pricing is a business-model innovation as significant as the technical claims — expect fast-follow across the agent market.
### 🧠 Learning Styles Assessment
- Reading level estimate: Flesch-Kincaid grade ~12–13 (long compound sentences, business and AI jargon: "tokenomics," "context engine," "prior authorization").
#### 👀 Visual learners
- Present: two images (title card, one product graphic).
- Improve: an architecture diagram of context engine + planner + Agent OS; a timeline visual of a weeks-long referral journey.
#### 🔊 Aural learners
- Present: nothing — text and images only.
- Improve: a 3-minute narrated walkthrough of the healthcare referral example.
#### 🗣️ Verbal learners
- Present: strong analogies ("digital front door") and a memorable antithesis (experiences vs. outcomes).
- Improve: define "long-horizon," "context engine," and "tokenomics" inline for readers outside the AI bubble.
#### 🧩 Kinesthetic learners
- Present: nothing interactive.
- Improve: a sandbox demo where readers watch an agent plan between simulated engagements.
#### 🧠 Logical learners
- Present: a numbered two-question framing (differentiation, tokenomics) mapped to answers.
- Improve: any quantitative evidence — completion rates, time-to-outcome, cost comparisons.
#### 🧑🤝🧑 Social learners
- Present: thread format invites replies.
- Improve: customer voices — the design partner speaks only through Taylor.
#### 🧘♀️ Solitary learners
- Present: self-contained essay, linkable product page.
- Improve: a "questions to ask your own org" checklist for evaluating outcome-driven agents.
#### 🌿 Naturalistic learners
- Present: the moat/compounding metaphor is loosely ecological.
- Improve: address compute-cost and energy implications of agents that run for months.
### 📊 Accessibility Assessment
#### Screen reader semantics
- Present: images carry alt text — but one is just "Image," and the other repeats the title.
- Improve: descriptive alt text conveying what the product graphic actually shows (WCAG 1.1.1).
#### Keyboard navigation
- Present: standard X/Reader rendering; links are focusable.
- Improve: platform-controlled; no content-level traps.
#### Contrast and readability
- Present: plain text, platform-styled — contrast is host-controlled.
- Improve: image-embedded text (product graphic) may fail contrast and isn't machine-readable (WCAG 1.4.5 images of text).
#### Text resizing and reflow
- Present: plain paragraphs reflow fine at 200%.
- Improve: none needed at content level.
#### Media and captions
- Present: static images only; no audio/video requiring captions.
- Improve: if the product graphic encodes a workflow, publish it as structured text too.
#### Forms and interactions
- Present: not applicable to this source — no forms.
#### Standards reference
- Gaps map to WCAG 2.2 AA 1.1.1 (non-text content) and 1.4.5 (images of text).
### 🌍 DEIB Assessment
#### Cultural breadth
- Present: Santander (global bank) among named customers; otherwise US-centric examples (Fortune 50, US healthcare, test drives).
- Improve: non-US outcome examples; note regulatory variance across regions.
#### Inclusive language
- Present: gender-neutral throughout; no idioms that block translation.
- Improve: heavy insider jargon ("tokenomics," "moats") excludes non-tech business readers; a plain-language recap would widen reach.
#### Neurodiversity support
- Present: clear headings, short sections, numbered list for the two-question framing.
- Improve: front-load a TL;DR; the outcome-pricing punchline arrives last.
#### Representation
- Present: two named voices, both male executives (Taylor, Bavor; Nadella cited).
- Improve: include design-partner practitioners — the people running these deployments.
#### Psychological safety
- Present: aspirational, non-hostile framing.
- Improve: acknowledge customer-side concerns (being pursued by a sales agent for months) rather than only business-side wins.
### ✅ Logical Validity
#### Strong reasoning
- **Causal gap analysis (Why Horizon section):** the argument that single conversations can't produce multi-step outcomes, illustrated with the concrete referral-scheduling case (dozens of texts and calls across three parties), is a well-constructed inductive argument from a real workload.
- **Problem-solution mapping (AI Tokenomics section):** the two executive concerns (differentiation, token spend) are stated crisply and each mapped to a specific product mechanism — clear argumentative structure, even if self-serving.
#### Fallacies
- **Appeal to popularity (bandwagon):** "hundreds of companies… including almost half of the Fortune 50" — adoption breadth is evidence of sales success, not that Horizon's new claims are true. Fix: cite outcome metrics from deployments.
- **Appeal to authority:** citing Satya Nadella's "cede value" article to validate the moat framing — borrowed prestige, not evidence. Fix: show a case where retained context measurably beat a competitor using the same frontier model.
- **Unfalsifiable assertion (begging the question):** "becomes a durable, expanding moat… that compounds" assumes the moat mechanism it needs to prove; no churn, retention, or performance data is offered. Fix: publish the compounding curve.
- **False dichotomy (mild):** "you don't pay for tokens, you pay for business outcomes" implies these are the only two models; hybrid and capacity pricing exist. Fix: position outcome pricing against the actual pricing landscape.
#### Tone of voice
- Confident, visionary, executive-facing marketing — "the most significant expansion," "durable, defensible moat," "we're excited to partner with you." Declarative certainty with no hedging anywhere in the thread.
### 🛡️ DARVO Analysis
- Not applicable to this source — a product announcement with no accusations, disputes, or victim/offender dynamics to reverse. No deny/attack/reverse patterns present.
## 🎯 Conclusion
- Bottom line: Horizon stakes out the next competitive frontier for enterprise agents — goal pursuit across months, powered by a context engine and long-horizon planning — and backs it with a genuinely notable pricing shift from tokens to outcomes. The technical claims are plausible but entirely unevidenced in this thread; the pricing commitment is the part you can hold them to.
- Next steps: read the Horizon product page for architectural detail; watch for independent attribution data from design partners; track whether outcome pricing propagates across the agent market within two quarters.
## 🔄 Refactor Appendix — The Next Horizon in Agents
> Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4).
> Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms
### 📚 Literature Atoms (H3)
#### [H1] Horizon shifts agents from conversations to long-horizon outcomes
> [!quote]
> Where agents to date have delivered customer experiences, Horizon delivers outcomes.
**Context & use**:
- Closing line of the "Introducing Horizon" section — the thread's thesis in one sentence.
- The experiences→outcomes antithesis is the cleanest handle for how the agent category is being redefined.
**Tags**: #xref #aphorisms
<!-- atomicity: 1 idea ✓ -->
#### [H2] Context engine plus long-horizon planning fills the platform gap
> [!quote]
> What existing AI platforms are missing is the context engine to stitch together those interactions, and long-horizon planning
**Context & use**:
- "Why Horizon" section, healthcare referral example.
- Names the two specific architectural components; useful vocabulary for evaluating any agent platform claim.
**Tags**: #xref #articleseed
<!-- atomicity: 1 idea ✓ -->
#### [H3] Customer-interaction memory is framed as a compounding moat
> [!quote]
> the memory of your customer interactions – becomes a durable, expanding moat for your business
**Context & use**:
- Third paragraph, learning-loop framing; restated in the Tokenomics section as the answer to Nadella's "ceding value" worry.
- The thread's most strategically loaded and least evidenced claim — flag as asserted, not shown.
**Tags**: #xref
<!-- atomicity: 1 idea ✓ -->
#### [H4] Sierra prices Horizon on outcomes, not tokens
> [!quote]
> you don't pay for tokens, you pay for business outcomes delivered
**Context & use**:
- Final section; Sierra explicitly absorbs token-spend risk.
- The one falsifiable commitment in the thread; the business-model innovation may outlast the technical framing.
**Tags**: #xref #articleseed #share
<!-- atomicity: 1 idea ✓ -->
### 🧠 Permanent Candidates (H4)
#### [P1] Outcome-based pricing transfers AI cost risk from buyer to vendor
**Claim** (one sentence, Courtney's voice): Outcome-based pricing moves token-spend risk onto the vendor, which reshapes both the buyer's procurement math and the vendor's agent behavior.
**Evidence pointer**: [[#H4]]
**Scope & conditions**: Holds when outcomes are contractually definable and attributable; breaks when attribution is disputed or outcomes are gamed (over-attribution inflates vendor revenue).
**Links**:
- [[AI pricing models]] — because extends
- [[Metered cloud billing]] — because contrasts on who bears variable cost
- [[WordPress AI plugin monetization]] — because applies-in ecosystem pricing debates
<!-- atomicity: 1 idea ✓ -->
### ✏️ Fleeting Queue
- [F1] Does "your context remains yours" appear in Sierra's actual contract terms? — https://sierra.ai/product/horizon #fleeting #inbox
- [F2] Who arbitrates a disputed outcome under outcome pricing? Open question for any outcome-priced AI contract. #fleeting #inbox
### → Things 3 (max 2 actions)
- [ ] Skim Sierra's Horizon product page for attribution/pricing mechanics — https://sierra.ai/product/horizon #things
<!-- deep-read:sections end -->
📄 **Full text:** [[Full Document Contents/Tweets/The Next Horizon in Agents.md|The Next Horizon in Agents by Bret Taylor]]
## Highlights
- With each interaction, your Horizon agent gets more intelligent. When your agent closes a sale (or gets rejected), it automatically learns to improve its decision-making for future customer conversations. The data foundation of Horizon – the memory of your customer interactions – becomes a durable, expanding moat for your business that compounds as your agents power more of your customer experience. ([View Highlight](https://read.readwise.io/read/01kxpftmd85rgb2tyb4jm9qjf8))
- Note: The moat claim — strategically loaded but entirely asserted; no churn, retention, or compounding data offered anywhere in the thread.
- Tags: atom/literature/h3
- Where agents to date have delivered customer experiences, Horizon delivers *outcomes*. ([View Highlight](https://read.readwise.io/read/01kxpftebc0q6tesb1422wt51f))
- Note: The thread's thesis in one sentence — the experiences→outcomes antithesis redefines the agent category.
- Tags: atom/literature/h1, aphorisms
- But an individual conversation is just a small ingredient in most complex business outcomes. One of our design partners for Horizon is a US healthcare provider using the platform to schedule referrals from primary care physicians to specialists. Getting those appointments scheduled can involve dozens of texts and phone calls – with the patient, with the specialist, and with the referring physician. What existing AI platforms are missing is the **context engine** to stitch together those interactions, and **long-horizon planning** to enable the agent to reason between engagements about what actions to take to maximize the odds of the appointment getting scheduled. ([View Highlight](https://read.readwise.io/read/01kxpfthjkvc434985hpz4vpde))
- Note: Names the two architectural components — context engine + long-horizon planning — grounded in the referral-scheduling example. Vocabulary for evaluating any agent platform claim.
- Tags: atom/literature/h2
- And with Horizon, you don’t pay for tokens, you pay for *business outcomes delivered*. We bear the burden of managing token spend, and you can focus on what matters most to your business: growing your customer base and driving revenue. ([View Highlight](https://read.readwise.io/read/01kxpftprkag5rh29ep56xncjv))
- Note: The one falsifiable commitment in the thread: outcome-based pricing shifts token-spend risk to the vendor. The business-model innovation may outlast the technical framing.
- Tags: action/h2, atom/literature/h1