Notion Rejects AI Lab Lock-In to Curb Surging Token Costs

Taming Token Town: Staying Flexible as Models Get Pricier With Notion — Ai4

Ai4

Trading model optionality for volume discounts destroys software margins because frontier labs are increasingly competing with

AI-assisted summary of Taming Token Town: Staying Flexible as Models Get Pricier With Notion — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 14, 2026, this talk features Sarah Sachs, AI Lead at Notion. According to the page description, the session explores how applied AI companies can navigate the volatile token market, manage compute procurement, and leverage model optionality and open-weight strategies as frontier labs raise prices and move up the stack to compete directly with their customers.

## TL;DR Sarah Sachs breaks down the economic and architectural realities of building scaled products in "Token Town," warning builders against vendor lock-in, stealth token inflation, and over-relying on expensive frontier GPUs for deterministic tasks. She outlines Notion's multi-model playbook, which utilizes dynamic routing along the Pareto frontier, trace-level benchmarking, open-weight transitions, and lightweight CPU-based TypeScript workers. Sachs also highlights critical security boundaries and demonstrates how multi-agent orchestration transforms both user workspaces and Notion's own internal development workflows.

## Atmosphere & style Sachs delivers a pragmatic, fast-paced, and witty presentation that blends insider engineering rigor with sharp economic realism. Embracing her moniker as the "Anna Wintour of AI models," she engages the room through candid polling, sharp industry critique, and transparent architectural playbooks. The session feels refreshingly grounded and empowering, cutting through marketing hype to offer tactical survival strategies for applied AI leaders.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=4s) Managing AI at scale is fundamentally an engineering and procurement challenge rather than just a financial modeling exercise. - [01:22](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=82s) 88% of enterprises are trapped in the AI assistant phase due to fragmented data silos, broken systems of record, and unpredictable economics rather than model intelligence limitations. - [03:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=207s) Token buyers face hidden inflation through models that consume several times more internal reasoning tokens for the same output while keeping sticker prices flat. - [03:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=207s) Trading away architectural optionality for upfront vendor discounts creates dangerous lock-in with suppliers that increasingly compete as first-party application builders. - [05:50](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=350s) Hotelling's law leads foundation model labs to cluster their pricing and capabilities just below the frontier leader, forcing moderate business tasks onto expensive near-frontier compute. - [08:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=507s) Dynamic routing along the cost-to-capability Pareto frontier protects businesses and end-users from model volatility and price spikes. - [11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Model benchmarking must evaluate the entire agentic trace—including tool calls and database lookups—rather than relying on marketing claims or isolated per-token metrics. - [11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Maintaining agility requires rebuilding evaluation harnesses frequently and never sacrificing architectural flexibility for short-term volume discounts. - [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) The performance gap between closed frontier models and open-weight alternatives tends to close within roughly three months, creating rapid cost-reduction opportunities. - [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) Applied AI relies excessively on expensive GPU tokens for deterministic workflows that are far better suited to lightweight CPU-based scripts and workers. - [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) Combining private data access, untrusted external content, and outbound network capabilities creates a 'lethal trifecta' that becomes more dangerous as reasoning models grow more capable. - [17:38](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1058s) Effective AI development is shifting toward multi-agent orchestration that chains specialized models together across shared workspaces. - [18:43](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1123s) Autonomous routing and remediation pipelines can drive major efficiency gains, handling roughly 40% of internal UI polish and bug tasks at Notion.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=4s) Introduction: The "Anna Wintour of AI Models" and the Engineering of AI Partnerships

Sarah Sachs opens by introducing herself as the lead of the AI engineering teams at Notion. Polling the audience on their familiarity with Notion to help frame her remarks, she finds about half the room raising their hands. While Notion is known as a friendly company and she considers herself a generally nice person, she emphasizes that managing Notion’s scale has turned AI into a serious, high-stakes operational domain. Much of her daily role involves sitting across the table negotiating AI vendor contracts, forging strategic partnerships with frontier model labs, and architecting how Notion delivers AI directly to its massive user base.

She jokes that the tech press recently dubbed her the "Anna Wintour of AI models"—a moniker she thoroughly enjoyed and immortalized with an AI-generated image. Grounding the presentation, Sachs stresses that navigating the AI market is fundamentally an engineering endeavor rather than a mere financial modeling exercise, made possible by her cross-functional team at Notion. Their work has focused on managing the rapid evolution and deployment of AI-enabled agents, both internal and external, within the Notion workspace.

[01:22](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=82s) The Evolution of AI in the Workplace and the 88% Enterprise Trap

Notion has historically operated as the universal system of record where teams organize docs, tasks, and meetings. However, Sachs notes that organizations adopt AI across three distinct evolutionary stages: 1. **AI as a Thought Partner:** The typical conversational chat interface (such as ChatGPT, Claude, or Gemini) where users ask questions in isolation and manually copy-paste responses into their work. 2. **AI as an Assistant:** The phase reached roughly a year prior, where AI actively operates within software environments—such as a Notion page—to automatically generate databases and execute direct user prompts. 3. **AI as Teammates:** Autonomous, permissioned agents capable of executing repetitive, multi-step workflows on their own as part of an integrated system.

When polling the room to see who feels their AI is truly operating as an integrated system, only a single audience member raises a hand (prompting Sachs to joke, "Come work for Notion"). Sachs highlights an industry study showing that 88% of enterprises remain stuck in the assistant phase, unable to advance to fully autonomous agentic workflows. She argues that this stall is caused not by limitations in model intelligence, but by fragmented data silos, broken systems of record, and unsustainable economics. Companies frequently find themselves using an overpowered, exorbitantly expensive frontier model simply to triage basic emails—a dynamic that is "completely overpowered, unnecessary, unwieldy, and sometimes dangerous." Beyond high costs, unpredictable token pricing creates a structural barrier that paralyzes organizational experimentation.

[03:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=207s) The Token Buyer’s Dilemma: Pricing Surges, Lock-In, and Supplier-Competitors

Sachs walks through the harsh economic realities of buying tokens at scale for Notion’s 100 million users. She outlines two common scenarios applied AI builders face: * **Exhibit A (Hidden Token Inflation):** A frontier lab upgrades a closed-weights reasoning model while keeping the per-token sticker price identical. However, the model silently consumes three times as many internal "thinking" and output tokens to complete the same task, tripling real execution costs. * **Exhibit B (Unjustified Price Hikes):** A provider releases a next-generation model that is 40% more expensive on sticker price alone. Sachs challenges the audience to name a single model upgrade across four generations that actually made their core product 40% better.

``` TRADITIONAL VENDOR LOCK-IN vs. MODEL OPTIONALITY

[ Frontier Lab / Competitor ] [ Applied Product Layer ] │ │ Aggressive Commitments Multi-Model Router │ │ Discount in Exchange for Evaluates: Quality / Trace Total Vendor Lock-in Routes: Closed & Open Weights ▼ ▼ [ Price Hikes & Margin Squeeze ] [ Cost & Quality Leverage ] ```

Compounding the problem, model suppliers are increasingly moving up the stack to compete directly with their applied AI customers. When enterprises commit to massive minimum-spend contracts in exchange for upfront discounts, they trade away their agility. If a supplier raises prices or shifts its focus, locked-in buyers have no exit. Furthermore, Sachs warns that companies acting merely as token-reselling wrappers without proprietary value will always lose to the upstream labs that produce the compute. While Fortune 500 giants hire armies of consultants and negotiate custom terms, the rest of the market lacks leverage. Sachs sees Notion's role as negotiating on behalf of the "Fortune 5 million," proving that in modern software engineering, "model optionality is the product."

[05:50](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=350s) The Cost-per-Capability Trade-off and Hotelling’s Law in AI

Sachs rejects the notion of a single "best" model, warning that reflexively chasing the newest frontier release is the fastest way to empty company pockets and experience an overnight 40% spike in Cost of Goods Sold (COGS). Citing research from Citadel, she describes an emerging bifurcation in the AI economy: simpler, lower-cost models efficiently handle high-volume routine tasks, while expensive frontier models are reserved strictly for complex orchestration and heavy reasoning.

To explain current model pricing behavior, Sachs invokes Hotelling's law—an economic principle illustrating why competing gas stations or beachside ice cream vendors cluster directly next to each other in physical space to capture the maximum share of the median market. In the AI landscape, foundation model labs mimic this behavior: if a lab does not possess the absolute frontier model, it positions its offering just a few cents cheaper than the market leader.

``` HOTELLING'S LAW IN AI PRICING

Cheap / Moderate Compute Frontier Lab Cluster ◄───────────────────────────────────────────────────────► [ Open-Weight / Specialized ] [ Model A ] [ Model B ] Underserved Moderate Tasks Clustered in the middle: High ROI for Routine Data Minor delta in capability and token pricing ```

This pricing dynamic creates a glaring market inefficiency: moderate, everyday business tasks are forced onto near-frontier models priced at near-frontier rates. Because benchmark leaderboards constantly shift, relying on a single closed-weights model guarantees that an application will become technologically obsolete or financially unviable half the time.

[08:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=507s) Operationalizing Model Optionality and the Pareto Frontier

Maintaining genuine optionality requires complete model agnosticism across external LLMs and specialized agents (such as Claude, Cursor, Decagon, and Codex). Notion introduced an "Auto" model selector where Notion’s routing engine dynamically assigns each query to the optimal model based on real-time cost-to-capability evaluations. Sachs clarifies that Auto is designed as a trust and efficiency mechanism for users rather than a high-margin upsell for Notion.

To illustrate why this matters, Sachs shares an anecdote from a trip to Australia, where she met the leadership of a biodegradable food packaging company. The company had diverted its entire internal IT team away from core business operations to manually test coding agents and evaluate model APIs. Sachs argues that forcing non-tech enterprises to become LLM infrastructure specialists is as broken as expecting every early dot-com business to employ specialized physical network engineers.

``` THE DYNAMIC PARETO FRONTIER

Quality ▲ ● Frontier Model (Complex Tasks) │ ╱ │ ●───────┘ (Optimal Balance) │ ╱ │ ●──────┘ (Open-Weight / Worker Task) │ ╱ │ ╱ ▲ Inefficient / Suboptimal Zone └────────────────────────────────────────► Cost ```

To solve this, Notion charts the Pareto frontier of AI performance, plotting real-world output quality against cost. By continuously riding this frontier, Notion automatically steers user requests toward the most cost-effective model that satisfies the task's required quality threshold, shielding users from volatile API pricing.

[11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Notion's Multi-Model Playbook and Trace-Level Evaluations

Sachs outlines the five core operational rules Notion uses to stay agile in the shifting token landscape: 1. **Build for Multi-Model from Day One:** Decouple application architectures from any single proprietary API. 2. **Evaluate on Real Value and Quality Trade-offs:** Measure end-to-end task performance rather than raw marketing claims. 3. **Switch Fast and Switch Often:** Notion rebuilds its evaluation harness roughly every two months and tunes its automated routing algorithms on a daily basis. 4. **Give Feedback Upstream:** Share structured evaluation data back with frontier labs to help improve their offerings. 5. **Never Trade Optionality for Discounts:** Long-term architectural flexibility consistently beats short-term volume discounts.

Sachs illustrates Notion’s evaluation framework using internal benchmarks published on social media. She highlights that standard marketing benchmarks (like per-token latency or price per isolated prompt) are useless for complex applications. Instead, Notion measures the total cost and latency across an entire agentic *trace*—including multi-step tool calls, web searches, and database operations. Because Notion possesses proprietary visibility into how users interact with workspaces, it can identify specific bottlenecks that public model leaderboards overlook.

[12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) Open-Weight Dynamics, CPUs over GPUs, and Notion Workers

Sachs emphasizes that open-weight models have reached a pivotal inflection point. While open-weight releases may not permanently surpass closed-weights frontier models, they steadily close the performance gap within roughly three months. Consequently, routine production traffic that requires a closed model today can often be shifted to a far cheaper open-weight model shortly after.

``` DETERMINISTIC WORKERS vs. MONOLITHIC LLM CALLS

[ User Action: Process Invoices / PDFs / Sync Tables ] │ ┌───────────────┴───────────────┐ ▼ ▼ [ Monolithic LLM ] [ Notion Workers ] - Expensive GPU Compute - Deterministic TypeScript - Non-deterministic output - Runs on lightweight CPUs - High latency - Fast API & Auth integrations ```

Sachs also presents the contrarian view that applied AI relies far too heavily on expensive GPUs. Many common enterprise tasks—such as parsing structured PDFs, transforming documents into CSVs, and executing scheduled data migrations—do not require a non-deterministic LLM to reinvent the process on every run. To address this, Notion introduced "Workers": lightweight TypeScript scripts executing directly on CPUs within Notion's infrastructure. Workers handle authentication, external API connections, and deterministic data manipulation quickly and cheaply. Backing this up with strong administrative governance, Notion provides visibility dashboards so enterprises can inspect exactly what their agents are doing, why specific models were routed, and what business value was produced.

- **Architecture over Model Choice:** System-level design, routing, and deterministic execution deliver more leverage than picking a single "winning" foundation model. - **Open-Weight Strategy:** Capitalize on the three-month capability catch-up window to migrate mature workflows away from premium closed APIs. - **CPUs for Deterministic Work:** Offload routine data formatting and API integrations to TypeScript workers rather than burning GPU tokens on probabilistic LLM loops.

[15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) Security's Lethal Trifecta and Multi-Agent Orchestration in Practice

Sachs warns of a major architectural risk she calls the "lethal trifecta" of AI security, which occurs when an application combines three elements in a single environment: 1. Access to sensitive private organizational data. 2. Exposure to untrusted, external user content. 3. The autonomous ability to communicate externally via APIs or network calls.

``` THE LETHAL TRIFECTA

[ Private Enterprise Data ] ▲ │ ┌────────────┴────────────┐ ▼ ▼ [ Untrusted Content ] ◄─────► [ External Comms ] ```

Counterintuitively, Notion discovered that more capable models pose greater security risks because their advanced reasoning makes them more adept at executing sophisticated prompt injections and jailbreaks. Rather than deploying a monolithic, all-powerful autonomous agent, Sachs advocates for compartmentalizing tasks, establishing sandboxed execution environments, and enforcing strict permission boundaries.

Transitioning to a live demonstration, Sachs enters what she cheerfully terms "sales mode," noting that multi-agent orchestration is where Notion delivers its greatest value. Instead of forcing human workers to act as full-time glue across disconnected tools, Notion provides an enriched, collaborative markdown canvas. In this environment, users can orchestrate specialized agents alongside human teammates: invoking Claude to analyze product copy, pulling in Decagon to surface real-time customer support trends, and looping in engineering leads to confirm technical constraints directly within the shared document.

[17:38](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1058s) Multi-Agent Orchestration and the Evolution of Workspace Collaboration

Modern AI workflows increasingly rely on chaining specialized models together rather than relying on a single monolithic system. In a concrete multi-agent development workflow, Claude can be instructed to continue building out a feature by synthesizing customer product feedback alongside technical discussions from engineering leads. Once Claude generates a pull request, the handoff continues by bringing in Codex to perform an automated code review and push iterative fixes directly to the branch. This multi-agent orchestration showcases how distinct models excel when assigned complementary roles in an end-to-end task.

```text +---------------------+ +--------------------+ +--------------------+ | Customer Feedback & | --> | Claude Agent | --> | Codex Agent | | Tech Lead Guidance | | (Implements PR) | | (Code Review & Fix)| +---------------------+ +--------------------+ +--------------------+ ```

While Notion initially launched as a collaborative workspace built exclusively for humans, its core mission has expanded to encompass all forms of collaboration: human-to-human, human-to-agent, and agent-to-agent. As AI handles an increasing share of modern knowledge work, the majority of collaboration occurring inside Notion will eventually take place directly between autonomous agents. Building for this future requires investing in flexible frameworks and maintaining maximum model optionality to ensure the best possible architecture for customers.

[18:43](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1123s) Internal Automation at Notion and Enterprise Impact

To illustrate agentic workflows in production, Sachs highlights Notion's internal Slack infrastructure. When employees drop bug reports or suggestions into a shared feedback channel, a self-learning custom routing agent analyzes the input, directs it to the appropriate team for passive awareness, and triggers a Claude-managed agent to implement the fix directly in the codebase. This automated pipeline handles approximately 40% of all polish tasks at Notion, removing tedious UI issues—such as overlapping buttons—from human backlogs so engineers can focus strictly on work at the "top of their license."

- **Slack Integration:** Incoming feedback in internal channels is autonomously parsed and routed by a self-learning custom agent. - **Automated Remediation:** A Claude-managed agent writes and implements code fixes for UI and polish issues without requiring manual developer intervention. - **Developer Productivity:** Approximately 40% of Notion's internal polish tasks are handled entirely through this automated agent workflow. - **Enterprise Adoption:** Marquee customer Vercel deployed Notion's agentic systems across their operational lifecycle—from feature intake to launch prep and releases—generating substantial ROI within their first two months of adoption.

[19:36](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1176s) Surviving Token Town and Buyer Advocacy

The AI compute and API landscape remains an unpredictable frontier, which is why the ecosystem is aptly nicknamed "Token Town." Because the token market is still nascent, pricing power is heavily skewed toward model providers, creating an imbalanced and often unfavorable environment for downstream buyers. To build sustainable products, engineering and product leaders must stay well-informed about model economics and actively advocate for model optionality to protect their customers, investors, and teams.

- **Market Dynamics:** The token economy currently functions like the "Wild West," characterized by rapid pricing shifts, shifting supplier dynamics, and market asymmetry against buyers. - **Community Engagement:** Sachs invites builders to connect on X/Twitter, LinkedIn, or email to discuss model orchestration, system architecture, and procurement strategies.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=4s) Anna Wintour of AI models — Sarah Sachs

> [01:22](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=82s) Come work for Notion — Sarah Sachs

> [01:22](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=82s) completely overpowered, unnecessary, unwieldy, and sometimes dangerous. — Sarah Sachs

> [03:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=207s) model optionality is the product. — Sarah Sachs

> [03:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=207s) Fortune 5 million — Sarah Sachs

> [11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Build for Multi-Model from Day One — Sarah Sachs

> [11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Switch Fast and Switch Often — Sarah Sachs

> [11:03](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=663s) Never Trade Optionality for Discounts — Sarah Sachs

> [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) lethal trifecta — Sarah Sachs

> [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) sales mode — Sarah Sachs

> [18:43](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1123s) top of their license. — Sarah Sachs

> [19:36](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1176s) Token Town — Sarah Sachs

## Mentioned Sarah Sachs, Notion, ChatGPT, Claude, Gemini, Citadel, Hotelling's law, Cursor, Decagon, Codex, Notion Workers, TypeScript, Slack, Vercel, X/Twitter, LinkedIn

## Video description Sarah Sachs - AI Lead, Notion

The token market is the Wild West, and if you’re building on frontier models, your pricing and margins can change overnight without warning.

As frontier labs raise prices, deprecate “good enough” models, and move up the stack into first-party products, applied AI companies are increasingly buying from suppliers who may also become their competitors. That makes compute procurement, model optionality, and open-weight strategy not just an ops problem, but a core product strategy decision.

[Source](https://www.youtube.com/watch?v=EV-tyA4RhZE)

中文简报

Notion拒绝AI实验室绑定以遏制飙升的Token成本 以模型选择权换取批量折扣会侵蚀软件利润率，因为前沿实验室正日益参与竞争 Sarah Sachs深入剖析了在“Token Town”构建规模化产品的经济与架构现实，告诫开发者警惕供应商锁定、隐性Token通胀以及在确定性任务上过度依赖昂贵前沿GPU的问题。她阐明了Notion的多模型策略，该策略结合了沿帕累托前沿的动态路由、Trace级基准测试、向开放权重模型的迁移，以及基于CPU的轻量级TypeScript工作节点。Sachs还强调了关键的安全边界，并展示了多智能体编排如何重塑用户工作空间以及Notion自身的内部开发工作流。 规模化管理AI本质上是一项工程与采购层面的挑战，而不仅仅是财务建模工作。 (0:04) 88%的企业因零散的数据孤岛、脱节的记录系统以及不可预测的经济成本而受困于AI助手阶段，而非受限于模型智能水平。 (1:22) Token采购方正面临隐性通胀：模型在标价保持不变的情况下，为实现相同输出会消耗数倍的内部推理Token。 (3:27) 放弃架构选择权以换取前期供应商折扣，会导致与日益作为第一方应用构建者参与竞争的供应商之间形成危险的锁定。 (3:27) 霍特林法则导致基础模型实验室将其定价和能力扎堆在仅次于前沿领跑者的区间，迫使中等复杂度的商业任务不得不使用昂贵的近前沿算力。 (5:50)

Reading summary

Sarah Sachs, head of AI engineering at Notion, argues that enterprise viability in applied AI depends on building multi-model optionality and system-level architecture rather than committing to single-vendor frontier models with volatile economics. The Economic Traps of Token Pricing (3:27) Sachs cites industry data indicating that 88% of enterprises remain stalled in the assistant phase of AI adoption, unable to reach autonomous workflows due to broken data systems and unsustainable unit economics. Applied AI builders face hidden token inflation, where closed reasoning models consume three times as many internal thinking tokens for the same task without changing the sticker price, or vendor price increases of 40% across model generations that yield marginal real-world improvements. Compounding this margin squeeze, model providers increasingly compete directly with their application-layer customers. Invoking Hotelling's law of spatial competition, Sachs observes that foundation model labs cluster just beneath the frontier market leader in both capability and price. This dynamic leaves standard business tasks burdened by near-frontier rates unless applications decouple from single proprietary providers. Dynamic Pareto Routing and Trace-Level Evaluations (8:27) To maintain architectural independence, Notion implements an automated multi-model routing engine that steers queries along a dynamic Pareto frontier of real-time cost versus task quality. Sachs emphasizes that public benchmarks and per-token sticker prices are misleading for complex applications. Instead, Notion evaluates models across entire agentic traces, incorporating multi-step tool calls, database interactions, and web lookups. Notion rebuilds its internal evaluation harness roughly every two months and tunes routing algorithms daily. Sachs advises builders to never exchange architectural agility for upfront volume discounts from model providers, arguing that long-term flexibility consistently outperforms locked-in pricing terms. Workload Partitioning with Open Weights and CPUs (12:23) Sachs highlights that open-weight models reliably narrow the capability gap with closed frontier models within approximately three months. This predictable catch-up allows engineering teams to systematically migrate mature, routine production traffic away from expensive proprietary APIs toward lower-cost models. Simultaneously, Sachs contends that applied AI overuses expensive GPU clusters for deterministic execution. Standard operations like parsing structured PDFs, converting documents to CSV format, and handling API authentications do not require probabilistic inference. Notion diverts these processes to lightweight TypeScript scripts called Workers that run directly on CPUs, reserving costly GPU tokens strictly for orchestration and complex reasoning. Security Boundaries and Multi-Agent Orchestration (15:07) Deploying autonomous agents exposes architectures to what Sachs terms the lethal trifecta: the intersection of access to private enterprise data, ingestion of untrusted external content, and the authority to execute external network calls. Because models with stronger reasoning capabilities are inherently more adept at executing prompt injections, Notion advocates for sandboxing execution environments and compartmentalizing tasks rather than deploying monolithic agents. In production, this compartmentalization functions through chained multi-agent orchestration. For example, Claude drafts pull requests from customer feedback and engineering guidance, while Codex performs automated code review and applies fixes. Sachs notes that Notion's internal Slack routing and automated coding pipelines already resolve approximately 40% of internal UI polish issues without manual engineering intervention. While Sachs anticipates that most future workspace collaboration will occur autonomously between agents, downstream buyers remain disadvantaged in an asymmetric token market, making architectural agnosticism and rigorous trace evaluations necessary defenses against vendor lock-in.

精炼阅读摘要

Notion 的 AI 工程负责人 Sarah Sachs 指出，在模型供应商频繁涨价、存在隐性消耗与锁定风险的“代币市场（Token Town）”中，应用层企业应将“模型灵活性（Model Optionality）”视为核心架构战略，通过多模型动态路由、确定性计算卸载与多智能体分权治理，打破大模型落地的成本与安全困境。 企业停滞困境与模型溢价陷阱 (1:22) 企业在利用 AI 时通常经历“思考伙伴”、“助手”到“自主队友”三个演化阶段，但行业研究显示高达 88% 的企业停留在助手阶段。Sachs 指出，这一瓶颈的核心症结并非模型智力不足，而是数据孤岛、破碎的系统记录以及不可持续的模型采购经济学。许多企业盲目调用价格昂贵、能力过剩的前沿大模型处理基础邮件分发等日常任务，造成了高昂且危险的资源浪费。 在代币采购层面，应用构建者面临严重的经济陷阱：一方面，部分前沿实验室在闭源推理模型标价不变的情况下，暗中增加三倍的内部“思考”代币，使实际运行成本翻倍增长；另一方面，新一代模型经常出现单价直接上涨 40% 的情况，却无法为终端核心产品带来等同比例的体验改进。受经济学中的“霍特林法则（Hotelling's Law）”影响，模型实验室大多在距离市场领头羊极微小的价格区间内扎堆定价，导致处理日常中等复杂度任务的高性价比模型严重供给不足。 基于帕累托前沿与Trace的动态路由 (8:27) 为对抗模型供应商向下游扩张带来的挤压与锁定风险，Notion 主张从第一天起就构建多模型架构，并明确反对为了短期的大额折扣而签署排他或高保底合同。Sachs 认为，没有专有数据和系统价值的纯模型套壳应用终将被上游算力厂商淘汰，应用层真正的护城河在于为缺乏议价能力的广大企业提供灵活的调度能力。 Notion 为此开发了“Auto”模型选择引擎，该引擎将各模型的实际任务质量与成本投射至动态的“帕累托前沿（Pareto Frontier）”上，在满足任务质量阈值的前提下，自动将用户请求导向成本最低的模型。在评测体系上，Notion 摒弃了孤立测量单个 Prompt 延迟或单价的传统基准，转而针对涵盖多步工具调用、网络检索和数据库操作的完整智能体追踪（Trace）进行端到端测算。Notion 内部大约每两个月重构一次评测基线，并每日对路由调度算法进行微调优化。 开源演进与CPU确定性计算架构 (12:23) Sachs 观察到开源权重模型的演进规律：开源模型通常会在约三个月的时间窗口内迅速追平前沿闭源模型的性能水准。这意味着企业可以在前沿闭源模型上验证新工作流，并在数月后将成熟的生产流量无缝切换至成本显著更低的开源模型，从而持续优化毛利结构。 针对行业对昂贵 GPU 算力的过度依赖，Sachs 提出了反共识的工程解法：企业内部大量高频任务（例如解析结构化 PDF、文档转 CSV 以及按计划同步表格数据）本质上是确定性逻辑，无须每次都调用昂贵且具备随机性的 GPU 大语言模型。Notion 为此设计了运行在 CPU 上的轻量级 TypeScript“Workers”，专门以毫秒级响应和极低成本处理认证、API 握手和数据确定性清洗，配合管理仪表盘监控智能体行为与业务产出，有效遏制了销货成本（COGS）的剧烈波动。 致命三元组防范与多智能体实战 (15:07) 在多智能体系统构建中，Sachs 警告必须高度警惕 AI 安全领域的“致命三元组”，即单一系统在同一环境下同时集成了“访问企业内部敏感数据”、“暴露于外部不可信内容”以及“具备自主对外发起网络或 API 调用”三项能力。测试表明，由于高级模型具备更复杂的推理能力，它们在面对精密提示词注入（Prompt Injection）和越狱攻击时反而更加脆弱，因此系统必须采用严格的沙箱运行、任务拆分与分权边界。 在协同作业模式上，Notion 转向多智能体链式编排以替代单体模型全包：在软件研发场景中，系统先由 Claude Agent 综合客户需求与架构规范生成拉取请求（PR），随后自动移交给 Codex 专用模型进行代码审查并推送增量修复。在 Notion 内部，基于 Slack 的自定义路由与智能体自动化流水线已能独立承担约 40% 的 UI 细节优化任务；企业客户 Vercel 也在部署该多智能体框架后，在两个月内跨越功能准入到发布运维的全流程实现了可观的投资回报。 Sachs 总结指出，当前的代币与 API 市场仍处于规则失衡、定价权极度偏向供给方的“狂野西部”阶段。随着未来的工作协作全面转向智能体与智能体之间的自动化交互，工程和产品决策者唯有保持架构解耦与模型灵活性，通过动态路由和精细化算力配置，才能在模型频繁迭代与价格波动的生态中维持长期的技术敏捷与商业可持续性。

Detailed notes — zh

## Background 该内容由频道 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台上，主讲人为 Notion 的 AI 工程负责人 Sarah Sachs。视频简介指出，当前的 Token 市场犹如狂野西部，随着前沿实验室提价并向应用层延伸竞争，算力采购、模型选择权与开源权重策略已不仅是运维问题，而是决定企业利润与生存的核心产品战略。 ## TL;DR 面对大模型供应商隐性 Token 膨胀、保底协议锁定及中端任务定价失衡的现实，Notion 践行了“模型选择权即产品”的工程理念，通过 Auto 模式动态追踪帕累托前沿以实现性价比最优。Sarah Sachs 提出应善用开源权重模型的快速赶超红利，并结合确定性代码（Notion Workers）分担无需 GPU 的常规任务，避免无谓的算力浪费。同时，针对“致命三元组”安全威胁与多 Agent 协作混乱的困境，Notion 将自身打造为多 Agent 协同枢纽，在内部研发与标杆客户中实现了兼具安全、高人效与强投资回报率的落地实践。 ## Atmosphere & style 整场演讲风格干练犀利、极具实战感，Sarah Sachs 以兼具前沿架构师与商业谈判官的双重视角，毫无保留地拆解了 Token 采购中的博弈黑幕与落地陷阱。现场节奏明快且金句频出，既有对“88% 的人沦为 Agent 协调员”等行业乱象的幽默调侃，也有对“宁可放弃折扣也要坚守架构灵活性”的冷峻洞察。聆听体验既像是一场直击应用层痛点的行业内幕大揭秘，又是一堂逻辑严密、兼具商业智慧与工程美感的 AI 架构实操课。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=4s) 工作空间中的 AI 正在从“思考伙伴”与“智能助手”向“自主队友”演进，约 88% 的企业仍停留在助手阶段，其瓶颈在于缺乏统一的数据记录系统（System of Record）与协作底座，而非模型参数本身。 - [02:46](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=166s) 企业在算力采购中常陷入“思考 Token 隐性膨胀”与“版本迭代 40% 溢价”的陷阱；为了眼前折扣而签署巨额保底承诺（Minimum Commitments），往往会导致企业丧失切换供应商的主动权。 - [05:16](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=316s) 占市场绝大多数的中小企业面对模型厂商缺乏议价杠杆，在算力成本与模型能力快速分化的当下，“模型选择权（Model Optionality）”本身已成为应用层产品的核心竞争力。 - [06:55](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=415s) 受经济学“霍特林法则”影响，模型厂商定价倾向于紧贴最强模型，导致大量中度任务被迫承担前沿模型的溢价，缺乏性价比匹配的中间层方案。 - [06:55](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=415s) 全球最强基础模型的归属频繁轮换，企业如果深度绑定单一供应商，意味着在技术迭代周期中有半数时间都在为昂贵且过时的技术买单。 - [08:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=507s) Notion 践行“模型不可知论”，通过 Auto 智能路由模式动态追踪“帕累托前沿”，让系统自动根据任务类型在质量与成本之间寻找最优解。 - [10:26](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=626s) 企业应当坚决拒绝排他性模型折扣以保留随时切换底座的灵活性，同时应基于整条执行链路（Trace-level）而非单次请求的理论跑分来评估模型表现。 - [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) 开源权重模型具备明显的“3 个月赶超曲线”，在闭源前沿模型上跑通的成熟业务流量，通常在数月后即可平稳迁移到同等水平的开源模型以大幅降低运营成本。 - [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) 绝大多数 Agent 任务根本不需要 GPU，直接利用 CPU 运行 TypeScript（如 Notion Workers）等确定性代码和状态机，比每次用 LLM 重新推理更高效、廉价且稳定。 - [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) Agent 落地必须防范由“私有数据、不可信内容、外部通信能力”构成的安全“致命三元组”，应通过任务隔离与受控沙箱环境化解高能力模型带来的潜在破坏力。 - [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) 未来的协作工作流不仅服务于“人与 Agent”，更关键的是构建富上下文的协作中心，促成“Agent 与 Agent”之间的高频协同与自动闭环。 - [18:43](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1123s) Notion 内部已有约 40% 的打磨类工程任务由自学习路由和自动编程 Agent 自主完成，其目标不是取代开发者，而是让人类工程师聚焦于最高维度的核心能力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=4s) 从“AI 领域的安娜·温图尔”到 Notion 的智能协作演进 作为 Notion 的 AI 团队工程负责人，Sarah Sachs 的日常工作不仅涉及构建前沿的 AI 产品架构，还要承担大量与顶尖 AI 实验室的商务谈判，探索如何在工程落地与商业合作之间架起桥梁。媒体曾戏称她为“AI 模型领域的安娜·温图尔（Anna Wintour）”，这一称号准确勾勒出她在模型选型、采购决策与产品工程之间的把控力。Notion 拥有超过 1 亿用户，其核心价值一直在于作为团队与数字员工的“通用工作台”与“单一记录系统（System of Record）”，将任务、会议和文档汇聚在统一空间中。 ```mermaid flowchart LR accTitle: AI 在工作空间中的三个演进阶段 subgraph Phase1["第一阶段：思考伙伴"] A["Thought Partner<br/>问答与复制粘贴"] end subgraph Phase2["第二阶段：智能助手"] B["Assistant<br/>页面内建库与直接操作"] end subgraph Phase3["第三阶段：自主队友"] C["Teammates / Agents<br/>权限隔离与自主协同"] end A --> B --> C ``` 随着大语言模型技术的演进，工作空间中的 AI 交互形态经历了三个清晰的发展阶段： 1. **思考伙伴（Thought Partner）**：用户在独立的对话框（如 ChatGPT、Claude 或 Gemini）中提问，再将生成内容复制粘贴到工作流中； 2. **智能助手（Assistant）**：AI 开始直接嵌入 Notion 页面中，能够自主构建数据库并执行具体操作； 3. **自主队友（Teammates）**：具备严格权限控制的 Agent 可以独立运行，承接高度重复性的复杂任务。 然而调研数据显示，约 88% 的企业仍然被困在“助手阶段”，无法顺利迈向协同系统。导致这种落地停滞的根源并非底层模型能力的欠缺，而是由于企业内部数据孤岛、缺乏统一记录系统以及协作链条断裂所致。 - **核心观点**：企业级 AI 落地的阻碍不在于模型参数本身，而在于缺乏统一的数据记录系统与协作编排底座。 - **行业现状**：88% 的企业在跨越从“对话助手”到“自主 Agent 系统”的鸿沟时遭遇瓶颈。 --- ### [02:46](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=166s) 算力采购的残酷现实：暴涨的思考 Token 与“40% 溢价陷阱” 在实际落地场景中，盲目调用最强的前沿模型往往会导致严重的资源浪费。使用最昂贵的顶尖模型去处理日常邮件分类，就像用重型武器处理微小琐事一样“能力过剩、毫无必要、难以驾驭且潜藏风险”。模型调用成本的不受控与不可预测，已经成为阻碍个人与团队发挥创造力的结构性壁垒。 从采购方（Buyer）的实际视角来看，当前企业面临着严峻的模型定价陷阱： | 现象类型 | 典型场景 | 实际影响与商业矛盾 | |:--- |:--- |:--- | | **现象 A：隐性 Token 膨胀** | 推理模型版本升级，单 Token 名义单价保持不变 | 针对特定任务，模型生成的输出 Token 与思考 Token（Thinking Tokens）暴增至原来的 3 倍，实际成本翻倍 | | **现象 B：40% 阶梯溢价** | 基础模型版本迭代（如版本号递增），定价直接上涨 40% | 连续经历数代模型迭代后，没有哪一次升级能让产品实际体验产生与之对等的 40% 飞跃 | 更为棘手的是，当下的大模型供应商（Foundation Model Labs）往往同时也是应用层企业的直接竞争对手。实验室通常通过“承诺保底用量换取折扣（Minimum Commitments）”的方式锁定客户。许多企业为了眼前的低价而签署巨额保底协议，结果陷入深度的供应商锁定（Vendor Lock-in）。一旦模型单价调整或出现更具性价比的选择，受制于高昂的保底承诺，企业将完全丧失切换路径。若一家公司的产品仅仅是对 Token 进行低维度的二次打包转售，其商业模式在供应商下场竞争时将脆弱不堪。 --- ### [05:16](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=316s) 为“500 万中小企业”争取议价权：模型选择权即产品 面对算力市场的混乱生态，《财富》500 强企业通常拥有专属的 AI 采购与架构团队，甚至雇佣大型咨询机构来制定策略。然而，占市场绝大多数的中小企业（即“500 万非 500 强企业”）在面对模型厂商时完全缺乏议价杠杆，既拿不到大客户规模折扣，也无力承担高额试错成本。 ```mermaid flowchart TD accTitle: 每秒每能力成本权衡与流量分流 UserReq["用户请求流入 Notion"] --> Router{"智能流量分流评估"} subgraph Bifurcation["模型二元分化策略"] Router -->|"常规/中度任务"| CostEff["轻量/高性价比模型<br/>降低 COGS 支出"] Router -->|"前沿推理/复杂编排"| FrontierM["顶尖前沿模型<br/>保证任务上限"] end CostEff --> Output["高质量低成本交付"] FrontierM --> Output ``` 在这个背景下，Notion 将自身定位为这 500 万家企业的代理谈判者与技术把关人。Notion 的核心竞争力不在于从头预训练最强的基础模型，而在于打造能够灵活驾驭多模型的顶层产品机制——**“在当下这个时代，模型选择权（Model Optionality）本身就是核心产品”**。 根据 Citadel 等机构关于 AI 经济学的分析，算力市场正呈现出明显的二元分化（Bifurcation）： - **常规与中度任务**：应当交由经过专门优化、极具成本效益的轻量模型处理； - **前沿与复杂编排任务**：才需要调配昂贵的顶尖模型。 盲目迷信“最新最贵即最好”只会让企业的营业成本（COGS）在一夜之间飙升 40%。唯有根据“每秒每能力成本权衡（Cost per Capability per Second Trade-off）”对流量进行精细化分流，才能实现商业与工程的可持续性。 --- ### [06:55](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=415s) 霍特林法则与模型定价：为何中端任务成了被遗忘的牺牲品 从经济学中的“霍特林法则（Hotelling's Law）”可以透视当前模型厂商的定价博弈。就像海滩上卖冰淇淋的商贩或街道两旁的加油站，两家最初相距甚远的店铺为了争夺中间地带的客流，最终都会向中心靠拢，并排驻扎在最优地理位置。 ```mermaid flowchart LR accTitle: 霍特林法则下的模型定价趋同 subgraph Vendors["模型厂商博弈"] V1["顶尖前沿模型<br/>（高定价锚点）"] V2["竞品厂商策略<br/>（仅比最强模型便宜数美分）"] end Vendors --> Squeeze["挤压中端需求"] Squeeze --> Issue["中度任务被迫承担<br/>前沿模型的溢价"] ``` 目前的 AI 模型定价生态正处于这种博弈之中：非第一梯队的厂商为了获取最大利润，往往只将价格定得比最强模型便宜几美分，直接瞄准高净值客群。这种定价策略导致大量介于“极简任务”与“极致前沿”之间的中度任务，找不到价格合理匹配的模型方案。 此外，根据第三方中立基准（如 Artificial Analysis）的追踪，全球最强模型的王座每个月都在轮换。如果应用层企业将自身底层完全绑定在某一家特定的供应商或单一模型上，在技术快速迭代的周期中，意味着有一半的时间都在为过时且昂贵的技术买单。 --- ### [08:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=507s) 模型不可知论与 Auto 模式：重构帕累托前沿 为了打破模型锁定并化解定价风险，Notion 践行了彻底的“模型不可知论（Model Agnosticism）”。在生态架构上，不仅兼容外部多样化的模型与 Agent 工具（例如 Decagon、Cursor、Claude、Codex 等），更在产品内部推出了由工程团队驱动的“Auto 智能路由模式”。 在 Auto 模式下，绝大多数用户无需纠结底层技术参数，而是由 Notion 的工程编排系统根据帕累托前沿（Pareto Frontier）动态匹配最适模型： - **同等质量下寻求最低成本**； - **同等成本下榨取最高性能**。 在与澳大利亚一家生物可降解食品包装公司的交流中，该客户提到他们不得不将整个 IT 团队抽调出来，专门研究代码 Agent 的配置与底层大模型选型。这种现象反映了行业基础设施的错位：就像互联网诞生初期要求每家实体企业都必须自己雇佣底层网络工程师一样不合理。Notion 将这种多模型动态路由与成本控制作为核心服务交付给客户，使企业无需承担沉重的算法选型包袱，即可享受最优的性价比红利。 - **帕累托前沿应用**：通过动态路由实时追踪质量与成本的最佳平衡点，规避单点模型失效风险。 - **按需调用价值**：Auto 模式不以赚取模型差价为目的，而是作为信任载体，为客户提供专业的模型治理与成本控制方案。 ### [10:26](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=626s) 帕累托前沿、多模型策略与放弃模型折扣的必要性 在模型选择上，企业的目标始终是在相同成本下获得更高质量，或在相同质量下获得更低价格。Sarah Sachs 指出，市场上只有极少数模型能真正落在最佳的性价比边界上，而 Notion 的核心任务就是为客户驾驭这一“帕累托前沿”（Pareto Frontier）。由于评估数据变化极快，不同时期的定价和模型版本（例如前沿推理模型及各代 GPT 序列）都在动态调整，客户完全不需要自己去苦恼哪一个是最佳选择，Notion 会在后台自动完成最优路由。 针对如何应对模型市场的剧烈变动，Sarah Sachs 为现场观众总结了一套核心应对策略： ```mermaid flowchart TD accTitle: 多模型工程与评估策略流程图 subgraph S1["架构与评估"] A["构建多模型架构<br/>(Multi-Model Architecture)"] --> B["全链路评测<br/>(Trace/Action 级延迟与成本)"] end subgraph S2["高频动态迭代"] B --> C["测试框架每2个月更新一次<br/>(Eval Harness)"] C --> D["默认路由算法每日自动调优<br/>(Auto Routing Daily)"] end subgraph S3["生态与商业选择"] D --> E["回馈前沿伙伴 Eval 数据<br/>(促成双赢生态)"] D --> F["坚守多模型选择权<br/>(拒绝排他性折扣绑定)"] end ``` - **快速切换与高频调整**：Notion 大约每两个月就会重构并更新一次评估测试框架（Harness），而默认的自动路由算法更是每天都在动态微调。 - **向模型提供商回馈评测数据**：将 Notion 内部的评估结果以邮件或报告形式回馈给前沿模型合作伙伴，这不仅能帮助供应商改进模型，也让双方建立起更有价值的合作关系。 - **放弃模型折扣以换取选择权**：Sarah Sachs 强调，为了单家模型供应商的价格折扣而牺牲多模型灵活性（Optionality）是得不偿失的。企业应保持随时切换底座的能力。 - **端到端链路评估（Trace-level Eval）**：评测不能只看单次 Query 的表面标价或延迟，必须评估每个 Agent 动作（Agentic Action）与整条执行链路（Trace）的综合成本与延迟，以及外部服务（如 Web 搜索）带来的独特附加值。只有 Notion 自己最了解自身的工作流，脱离实际场景的理论跑分毫无意义。 --- ### [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) 开源权重模型的赶超曲线与确定性代码（Workers）的价值 谈到开源权重（Open-weight）模型，Sarah Sachs 认为这是目前整个行业最值得投入的领域，且已经成为一股不可阻挡的潮流。以 GLM-5.2 等模型为例，它们直接推高了业界的帕累托前沿曲线，在质量上展现出顶级 Agent 模型的表现，而调用成本却大幅降低。 ```mermaid flowchart LR accTitle: 开源权重模型对闭源模型的赶超曲线 subgraph T0["当前阶段 (Day 0)"] A["闭源前沿模型<br/>(Closed Frontier Models)"] -->|高成本承接关键业务| B["生产环境稳定运行"] end subgraph T3["3个月后 (Month 3)"] C["开源权重模型赶超<br/>(Open-weight Models)"] -->|平替原有闭源流量| D["显著降本 & 本地化掌控"] end B -.->|迁移成熟流量| C ``` Sarah 引用了 Philipp Schmid 的经典观点：虽然开源权重模型未必能在绝对能力上时刻领先闭源前沿模型，但它们总是能以极快的速度“赶上来”（Catch up）。如果某项业务流量在当前使用闭源模型已经跑通且表现良好，那么大约三个月后，这部分流量大概率就能平稳迁移到同等水平的开源权重模型上，从而大幅削减成本。 与此同时，Sarah 提出了一个更为务实的技术判断：**绝大部分 Agent 任务根本不需要 GPU，CPU 即可高效胜任**。在许多客户工作流中，常见任务往往是遍历邮件、解析 PDF 并转换为 CSV 表格再进行分发。如果让 LLM 每次都去重新“发明”这一套处理逻辑，不仅极其昂贵、速度缓慢，而且缺乏确定性。确定性代码和状态机（State Machines）在工程中具有不可替代的价值。 - **Notion Workers 架构**：Notion 推出了名为 Workers 的解决方案，直接在 Notion 内部运行 TypeScript 代码。这些代码既能访问内部与外部 API，也能处理自定义鉴权（Auth），实现任意数据与 Notion 之间的双向同步，同时执行高度确定性的业务逻辑。 - **治理与可见性（Governance & Visibility）**：企业需要让用户清晰看见 Agent 的执行步骤、决策依据以及创造的实际业务价值，确保系统具备高可维护性与控制力。 - **架构重于模型选择（Architecture over Model Choice）**：构建超越单一 Token 的产品价值，不盲目追求接入所有模型，而是围绕用户的帕累托曲线做最优解。 --- ### [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) 安全“致命三元组”与多 Agent 协作工作流 离开 Token 层面，真正决定企业应用成本与质量的还有安全架构与编排系统。Sarah Sachs 提出了 Agent 落地中的**“致命三元组”（Lethal Trifecta）**安全隐患： ```mermaid flowchart TD accTitle: 致命三元组安全威胁图 A["私有数据<br/>(Private Data)"] --- D{{"致命三元组<br/>(Lethal Trifecta)"}} B["不可信内容<br/>(Untrusted Content)"] --- D C["外部通信能力<br/>(External Communication)"] --- D D ==> E["巨大安全漏洞与数据泄露风险"] ``` 评测发现，**模型能力越强，潜在的攻击危害往往也越大**。因此，与其盲目依赖单一全能的庞大模型作为单体编排器（Monolithic Orchestrator），不如通过任务隔离（Task Segregation）、沙箱环境（Sandboxes）以及带受控浏览器访问的安全执行环境来化解风险。 在多 Agent 协同方面，现场引发了强烈共鸣：许多公司引入了 5 到 6 个不同的 Agent，结果员工的全部日常工作变成了充当协调员、确保这些 Agent 之间能正常配合。Sarah 笑称，现场大约 88% 的人深陷于这种泥潭，剩下的 12% 则在嘲笑大家。 Notion AI 将自身定位为多 Agent 协作中心（Multi-Agent Orchestration Hub），通过富上下文的协作空间将人类与各类专用 Agent 串联起来： ```mermaid flowchart LR accTitle: Notion AI 多 Agent 协作工作流 A["Notion 任务/上下文"] --> B["Claude Agent<br/>(生成初始 PR/方案)"] C["Decagon Agent<br/>(注入真实客户反馈)"] --> D["人类 Tech Lead<br/>(Dan: 提出架构批注)"] B --> E["上下文综合汇总<br/>(合并客户诉求与评审意见)"] D --> E C --> E E --> F["Codex Agent<br/>(代码审查与自动修复)"] ``` - **全流程协同而非干瘪的 Markdown**：在 Notion 页面内，Claude 负责思考并提出方案，Decagon 等客户声音 Agent 实时提供用户反馈，Tech Lead（如团队工程师 Dan）在页面中直接添加评论和架构指针，随后 Claude 综合所有信息更新方案，最后交给 Codex 审查 Pull Request 并提交修复补丁。 - **Agent 与 Agent 之间的协作**：Notion 最初是为人类协作而设计的，但在当下知识经济中，越来越多的工作由 Agent 承担。Notion 的愿景不仅是“人与人”、“人与 Agent”的协同，更是成为“Agent 与 Agent”之间高频交互的核心平台。 --- ### [18:43](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=1123s) Notion 内部实践成果与应对 Token 狂潮的倡议 在 Notion 内部，这套多 Agent 协作体系已经全面深入日常研发流程。Sarah Sachs 展示了 Notion 内部 Slack 频道的真实应用场景： - **自学习路由与自动修复**：团队在 Slack 中设立了一个名为 `#all-feedback` 的反馈频道。一旦有人抛出 Bug 或改进建议，Notion 自研的自学习路由 Agent 就会自动将任务派发给对应团队，在保持相关人员被动知情（Passively Aware）的同时，调度由 Claude 托管的 Agent 自动编写并上线修复方案。 - **释放高价值工程生产力**：Notion 内部约 40% 的打磨类任务（Polished Tasks，例如按钮重叠等 UI 细节修补）均由上述 Agent 自动化链路完成。这并非取代人类，而是让工程师将精力聚焦于自身能力的最顶峰（"Top of their license"），大幅提升整体人效。 - **标杆客户落地效果**：知名云基础设施平台 Vercel 作为标杆客户，已将从需求接入、发布筹备到正式 Release 的全生命周期放在 Notion AI 架构上运行，在短短数月内获得了显著的投资回报率（ROI）。 Sarah Sachs 在结语中呼吁，当前的 Token 市场犹如狂野西部，规则尚未成熟且对企业买方存在诸多不公与信息不对称。应用层开发者必须保持敏锐与知情，通过架构设计和多模型选择权为自身及客户争取最大利益。她在演讲最后公布了自己的联系方式（Twitter/X: `@serumsax`），欢迎业界同行就相关系统架构与工程问题展开交流探讨。 ## Notable quotes > [05:16](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=316s) 在当下这个时代，模型选择权（Model Optionality）本身就是核心产品 — Sarah Sachs > [02:46](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=166s) 使用最昂贵的顶尖模型去处理日常邮件分类，就像用重型武器处理微小琐事一样“能力过剩、毫无必要、难以驾驭且潜藏风险” — Sarah Sachs > [08:27](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=507s) 同等质量下寻求最低成本；同等成本下榨取最高性能 — Sarah Sachs > [10:26](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=626s) 为了单家模型供应商的价格折扣而牺牲多模型灵活性（Optionality）是得不偿失的。企业应保持随时切换底座的能力 — Sarah Sachs > [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) 虽然开源权重模型未必能在绝对能力上时刻领先闭源前沿模型，但它们总是能以极快的速度“赶上来” — Philipp Schmid > [12:23](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=743s) 绝大部分 Agent 任务根本不需要 GPU，CPU 即可高效胜任 — Sarah Sachs > [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) 模型能力越强，潜在的攻击危害往往也越大 — Sarah Sachs > [15:07](https://www.youtube.com/watch?v=EV-tyA4RhZE&t=907s) Notion 的愿景不仅是“人与人”、“人与 Agent”的协同，更是成为“Agent 与 Agent”之间高频交互的核心平台 — Sarah Sachs ## Mentioned Sarah Sachs, Notion, Anna Wintour, ChatGPT, Claude, Gemini, Citadel, Artificial Analysis, Decagon, Cursor, Codex, GLM-5.2, Philipp Schmid, TypeScript, Notion Workers, Slack, Vercel, Ai4 ## Video description Sarah Sachs - AI Lead, Notion The token market is the Wild West, and if you’re building on frontier models, your pricing and margins can change overnight without warning. As frontier labs raise prices, deprecate “good enough” models, and move up the stack into first-party products, applied AI companies are increasingly buying from suppliers who may also become their competitors. That makes compute procurement, model optionality, and open-weight strategy not just an ops problem, but a core product strategy decision. [Source](https://www.youtube.com/watch?v=EV-tyA4RhZE)

Source quotes

[5:16] Sarah Sachs: 在当下这个时代，模型选择权（Model Optionality）本身就是核心产品 [2:46] Sarah Sachs: 使用最昂贵的顶尖模型去处理日常邮件分类，就像用重型武器处理微小琐事一样“能力过剩、毫无必要、难以驾驭且潜藏风险” [8:27] Sarah Sachs: 同等质量下寻求最低成本；同等成本下榨取最高性能 [10:26] Sarah Sachs: 为了单家模型供应商的价格折扣而牺牲多模型灵活性（Optionality）是得不偿失的。企业应保持随时切换底座的能力 [12:23] Philipp Schmid: 虽然开源权重模型未必能在绝对能力上时刻领先闭源前沿模型，但它们总是能以极快的速度“赶上来” [12:23] Sarah Sachs: 绝大部分 Agent 任务根本不需要 GPU，CPU 即可高效胜任 [15:07] Sarah Sachs: 模型能力越强，潜在的攻击危害往往也越大 [15:07] Sarah Sachs: Notion 的愿景不仅是“人与人”、“人与 Agent”的协同，更是成为“Agent 与 Agent”之间高频交互的核心平台

Source quotes

[0:04] Sarah Sachs: Anna Wintour of AI models [1:22] Sarah Sachs: Come work for Notion [1:22] Sarah Sachs: completely overpowered, unnecessary, unwieldy, and sometimes dangerous. [3:27] Sarah Sachs: model optionality is the product. [3:27] Sarah Sachs: Fortune 5 million [11:03] Sarah Sachs: Build for Multi-Model from Day One [11:03] Sarah Sachs: Switch Fast and Switch Often [11:03] Sarah Sachs: Never Trade Optionality for Discounts [15:07] Sarah Sachs: lethal trifecta [15:07] Sarah Sachs: sales mode [18:43] Sarah Sachs: top of their license. [19:36] Sarah Sachs: Token Town

Original transcript

[0:04] I'm Sarah. I lead the AI teams at Notion. Raise  your hand if you're familiar with notion. No [0:07] pressure. It just helps me frame. Okay, about half  of you. Um, so I lead engineering at Notion um in [0:14] our AI space. This talk is about how to build  those AI products, but it's also how you can [0:18] set up a sustainable software factory and think  about building sustainably for your organization. [0:23] So, let's jump into it. Notion's a nice company.  I'm a generally nice person, but I also um it's [0:29] becoming a pretty serious product and a lot of  my job has to been negotiating our AI contracts [0:35] and doing a lot of specific work around thinking  about how we partner with the labs and bring AI [0:39] to all of our customers um as an engineering  lens but also a lot of this partnerships. [0:43] Um recently um I got described um by the press  as the Anna Winter of AI models which obviously [0:49] I loved. Um, so we little AI image of that you  know before also I get started I just want to [0:55] remind you that none of this work is ever done  alone and it's done with a team and especially [1:00] with what I'm going to describe a lot of it is an  engineering effort it's not a finance projection [1:05] effort um this is me with some of our team at  our launches and we've been pretty excited to [1:10] see how the growth of our product has changed  particularly with the rise of um AI enabled [1:16] agents inside and outside of notion this is our  growth inside of notion by orchestrating them. [1:22] So for those of you who know about notion,  again, notion's always been that universal tool [1:27] and that system of record for your team and  your agents. And it's the perfect place for them [1:31] all to work together. You can keep your tasks,  your meeting, your documents all in one place. [1:37] But we found over time that as AI has evolved and  entered our workspaces, not everyone is adopting [1:43] it the same way. AI can be a thought partner.  Um, think about it as like a chat GPT cla Gemini [1:50] thread. You ask questions and you copy paste it um  into your element. AI as an assistant is where we [1:56] got to I would say around last September. You  could actually start having AI operate within [2:01] for instance an ocean page creating databases  doing things for you. AI is teammates is what [2:06] we launched um almost a year ago now or I guess  six months ago um where you can do repetitive [2:12] work and actually have permission agents working  on their own. And you can imagine if these agents [2:16] work a system. Um, as a show of hands, how many  of you feel like your AI is working as a system? [2:24] Right. Okay. You in the back. That's awesome. Come  work for Notion. Um, most people don't feel that [2:29] way. We did a study. Um, we found about 88% of all  enterprises have trouble leaving that assistant [2:35] phase. Why is that, right? AI is so capable yet  none of us are able to use it to its capability. [2:41] Um my argument is that actually comes from  siloed data systems of record and collaboration [2:46] not from model capabilities and we see this  a lot. Cost is a huge example. Um the first [2:53] time I gave this talk in May these are just the  headlines from that week right and so you end up [2:57] being like this guy um trying to you know maybe  triage your emails um using the most capable [3:03] model out there. It's completely overpowered,  unnecessary, unyieldy and sometimes dangerous. [3:10] oftentimes beyond that it's not even the fact  that it's too expensive. It's simply a structural [3:15] barrier. It is so uncontrollably expensive  that it's hard to understand how to use it [3:21] in a way where individuals can kind  of wield it for their own creativity. [3:27] And I want to talk a little bit about  what that's like as a buyer. So, you know, [3:30] notion has 100 million customers and so we're  buying a them. Um let's look at exhibit A. Imagine [3:37] a reasoning model. A closed weight model gets  upgraded and it's the identical per token pricing. [3:43] However, it's using three times as many  output and thinking tokens for certain tasks. [3:48] Exhibit B, maybe there's a new model. Maybe it  goes from seven to eight or whatever that first, [3:53] you know, numerator is in their model naming  schemes and it is 40% more expensive. Is it making [3:59] your product 40% better? I think we've had about  four generation of model updates. Can anyone say [4:04] that one model update has made their product 40%  better? Okay, something is broken in the system. [4:11] On top of that, your supplier of models is  often also your competitor. Um, a lot of the [4:16] way that these contracts are structured, they're  structured with minimum commitments in exchange [4:20] for discounts. And often times enterprises  make the mistake of opting for larger discounts [4:25] and then putting themselves in an area of lock  in. If you tie yourself to what's cheap today, [4:30] your prices may change. something might be 40%  more expensive, but your commitment is so high, [4:35] you have no other option. I think this is a  pretty common narrative. The second version [4:40] is that your entire product is tokens. You're just  a token processing factory without any value ad. [4:46] You will always have a bad deal. It will  always be worse for you to buy, restructure, [4:50] and resell tokens than the competitor that sells  them to you. This is just principles of goods. [4:57] The Fortune 500 has dedicated teams to navigating  this. They have consultancy agencies. They have [5:02] organizations and they figure it out for  themselves. Everyone else is left without [5:06] leverage. If they don't have the ability to  negotiate commitments at such a high scale, then [5:11] they don't have the ability to have discounts that  create reasonable products for their customers. [5:16] At Notion, I view my job as negotiating for  the Fortune 5 million, for those non-fortune [5:21] 500 companies that want access to these products  and experiences where I believe optionality is [5:26] important. And the way that we do that is we don't  think about our job as training the best model. [5:33] Of course, there are instances where we invest  in the openweight community in reinforcement [5:36] learning, which we'll talk about, but also a lot  of it is building the best product that utilizes [5:41] many models. I would argue that in today's day  and age, model optionality is the product. Okay, [5:50] before we go further, I think there's a  framing that I want us to all agree on. [5:54] It's called the cost per capability per  second trade-off. Today a customer said [5:58] what's the best model that notion loves and I  said come to my talk that's no such thing okay [6:04] I think the problem is when you keep assuming  the next and best is the best for you [6:08] that's when your pockets empty that's when you  increase 40% overnight in cost of goods served [6:14] Citadel released this interesting white  paper about this which is not to say that [6:18] frontier AI intelligence is not important  it very much is but for the economy at large [6:24] there's a bifurcation of models Sometimes  simpler models are the more cost-effective, [6:29] but then for that frontier use or orchestration,  then the most capable models are the most [6:34] appropriate. And we're going to talk about how to  navigate that. Not all traffic is created equal. [6:40] I actually think in the past probably three weeks  with the Kim K3 release, people are starting to [6:44] realize this more than when I started talking  about this. But at notion, you can imagine we have [6:49] those moderate tasks and those frontier tasks,  and the way that we address them is differently. [6:55] I'm going to take a step back and go through  another kind of economic theory. Has it ever [6:58] noticed that there's always gas stations next  to each other? And you're like, how can that be? [7:02] How can that make sense? And how why is one more  expensive than the other? Why would they ever do [7:06] that? Okay, this is called hoteling's law. You can  also a really good example is I think like selling [7:12] ice cream at a beach. If you imagine pricing,  um, but you can also think about it as distance [7:17] geographically for a gas station. When you're  super far apart from each other, you're arguing [7:22] over the middle. Okay, so over time, this theory  says that more and more vendors move towards the [7:29] middle until eventually they end up right next  to each other because that's actually the optimal [7:33] location because then they can cover the most  confident amount of traffic in either direction. [7:39] I would argue that's how pricing works for AI  models today. I would argue that if you're not [7:44] the most capable model, then all you need to do  is be a couple cents cheaper, right? And and that [7:50] makes sense. There's no blame on that, but those  are the industry pressures on model pricing today. [7:57] And what's being left out with this pricing  structure is those kind of moderate tasks [8:01] that don't need the second most expensive  model pricing out there. And on top of that, [8:07] the best model is changing quickly. You can  see this is slightly old because I didn't [8:11] want to ruffle any feathers, but artificial  analysis is a great universal benchmark. [8:16] Um, you can see that the best model is always  changing. And so if you're locking yourself [8:20] into one specific model, you are locking  yourself into being obsolete half the time. [8:27] So optionality is your leverage. The only  way you have reasonable negotiating power [8:33] is by maintaining your product's capability to  have optionality. If you have optionality and [8:39] you know what's best for your customers  and that cost per capability trade-off, [8:43] then you're able to make educated points about  what your customers need, what they don't need, [8:47] and what the right backup options are for them.  [snorts] The first version of that is model [8:52] agnosticism, excuse me. Notion does  this not just um external AI models, [8:58] but also with um agents like Decagon,  Cursor, Serville, Claude, Codeex. [9:05] We also have in our product uh majority  of customers are choosing auto in which my [9:11] team in the engineering organization is actually  deciding what the best model is for our customers [9:17] with usage based pricing. This is actually a  product for our customers. It is not a margin [9:20] making product for us. It is a product where our  customers are trusting us to have the expertise [9:26] on what the right model is for their cost and for  what they need. You know, I was talking to a bunch [9:31] of customers in Australia, actually, and I was  talking with a really interesting biodegradable [9:36] food company, and he was describing to me how  he's rededicated his entire IT team to thinking [9:42] about setting up coding agent skills and model  selection. Why? Why would a biodegradable food [9:48] company be investing in that expertise? Because  they have to today. And I think that's a broken [9:52] system. It's like the internet was invented and  every company needs to have network engineers on [9:56] staff. That is not appropriate. Okay, so notion  is trying to sell that service to our customers. [10:03] In the auto model, it kind of depends. Who here  is familiar with the concept of a parto frontier? [10:09] Okay, some of you. If you look at um quality on  the left and cost on the right, you can think [10:15] about at any point in time there might be another  model that could be cheaper for the same quality [10:21] or better quality and more expensive, but  that's okay because it's better quality [10:26] or better quality for the same price, right? And  so there's only a certain subset of models and [10:31] this data is illustrative and old. So keep that in  mind. You'll see the there's not even Opus 5 yet [10:38] and the pricing has changed for the GPT56 models.  But you can understand that if we're doing our [10:43] job well, we're writing that Parto Frontier for  our customers, right? And we're never routing [10:49] them necessarily to something that isn't an  optimal choice, but they also don't need to [10:52] figure out what that optimal choice is. So you  guys love taking photos of slides. I've noticed [10:57] that this is that time. So go ahead, just take  it all out at once. Okay, this is the strategy. [11:03] You need to build from multimodel. Evaluate on  the value on that quality trade-off perspective. [11:09] Switch fast and switch often. We're changing our  harness about every two months and we're changing [11:14] our model in that default auto routing  daily. We're optimizing that algorithm [11:20] and give them something back. So, we're going to  talk about how we're sharing those evaluations [11:25] because those emails are actually helpful  to our Frontier partners. And I don't think [11:29] discounts are worth losing optionality. That is a  subjective perspective that I've tried to lay out [11:34] for you. Of course, that's the choice that you  can for your own enterprise as a stakeholder. [11:40] So, here's an example of an old eval that we put  on Twitter a couple months ago when evaluating. [11:46] You can see that it's not just cost  that we're looking at. We're looking at [11:50] what's unique value ad from the  web search provider that don't have [11:54] what's the cost per trace or per agentic action  not just per query what's the latency for the [11:59] entire trace not just the query right the only  people that can have expertise on this is us [12:05] because we are the only ones that know how  notion workflows work so if we were to look at a [12:09] a sticker price of latency or cost it would not be  valuable for us and because we've invested in that [12:16] eval technology we can then deliver the best price  to latency to cost trade-off to our customers. [12:23] We about open weight a little bit. I think this is  probably the biggest gap in industry can invest in [12:29] openweight models are strong enough to handle  this. I think this is actually just like very [12:33] invogue now. So maybe I don't need to sell  you on it. GLM52 I would say is probably [12:38] the biggest industry shift that we've  seen of bumping up that Pareto curve. [12:44] Look at its quality delta there, right? It  plays like a top class agentic model with [12:50] much more affordable use. Some of the newer ones  probably are similar range. Um, it's dynamic. [12:59] I love this slide from Phililly,  so I asked to steal it. Um, [13:03] open weight models will never necessarily be ahead  of closed models, but the point is that they catch [13:09] up. So if you're satisfied with traffic today that  you feel like is working well then in three months [13:14] it probably can be on an openweight model. That's  not to say that there's not traffic in two months [13:19] that you couldn't do today on any model that  you think is valuable on a closed model. Right? [13:25] And the last one is kind of a controversial  take. I actually don't think we need CPUs for [13:29] most of this work. GPUs, excuse me. I think CPUs  can do a lot of it. When we see our customers [13:34] having workflows that are, you know, going through  their emails, taking PDFs, turning them into CSVs, [13:39] and then sharing them again, why does an LLM  need to reinvent how to do that every time? [13:45] It's expensive. It's slow. It's not deterministic.  State machines have value. There's a reason that [13:51] deterministic code is valuable, right? Notions  released something called workers. This is our [13:57] solution to this. There's lots that I would  encourage you to explore in your own product [14:01] framework. The way that workers work is they're  typescript code that's actually served on notion. [14:07] That typescript code can access internal  and external APIs. You can create your own [14:12] off scenarios, anything that you need and you  can bring any data into notion and from notion [14:18] and on top of that we're able to  do deterministic logic as well. [14:21] This has been a significant cost reducer  and quality improver for our customers. [14:27] I would say the last one is governance. Um,  visibility is something that you need to have [14:32] in order to do this well. Maintainability and  control. If your customers are trusting you with [14:37] these decisions, you need to show them why you  made them. This is what it looks like in notion. [14:43] Not only can you see how you're  building and what your agents are doing, [14:46] you can also understand the value that they're  driving. In summary, before we leave token town, [14:53] architecture over model choice, I would say is the  number one thing that you should be investing in. [14:58] Open weight optionality doesn't mean you have  to offer every model. Think about what that [15:02] parto curve means for your customers and  build that value that transcends tokens. [15:07] I'm gonna leave token town for a  second. That's me departing token town. [15:13] There's a lot of other challenges that I  think have nothing to do with model selection [15:16] that really drive cost and quality. The first is  security. There's a concept of a lethal trifecta. [15:21] This is it. When you have private data, untrusted  content and the ability to communicate externally, [15:28] you are releasing your customers to a lot  of security vulnerabilities. And in fact, [15:33] we have found that the more capable the  model, the more capable the attack. Okay, [15:38] so in this case, maybe model capability and  autonomy is not what's best for your customers. [15:42] Maybe it's actually segregating the task  and not having a monolithic orchestrator. [15:48] Similarly, sandboxes and computers, safe execution  environments with access to browsers. That is not [15:54] necessarily a bigger model, but it will unlock  capability in a way that others might rely on [15:59] a big model to do so. And lastly, multi-agent  orchestration, which is where I'm going to end us. [16:06] How many of you feel like this  is your company's workflow? [16:09] It's like you're all here, you've  invested in six different agents, [16:12] and then it's your full-time job at work  to make sure that they all work together, [16:17] right? I mean, this is, by the way, um where about  88% of of everyone is, and those are probably the [16:24] ones of you that are laughing or you're the  12% that are looking down at us and laughing. [16:29] Okay, this is me in Vegas at AWS reinvent  telling Tain to tell everyone buy Notion AI. [16:34] For the next two minutes and 50 seconds, I  will be selling Notion AI. always be closing, [16:38] always be hiring, always be selling. So this is  how notion handles multi- aent orchestration. [16:42] I'm trying to make it not just a notion pitch, but  with me, this is me selling. So today in notion, [16:49] you can handle any task, analyze it, and  collaborate. Imagine if when going through [16:56] those specific tasks, you actually while looking  at it decide that you want Claude to take a stab. [17:03] Claude is thinking. It's asked a question.  Right? These are agents that are calling on [17:08] cloud agents orchestrated through notion. It's  not a markdown file. It's fully collaborative. [17:14] You can also pull in other agents like decagon,  your customer voice agent that can actually tell [17:18] you what your customers are looking for. Imagine  that we have this advice from DecaGon. And now we [17:25] need to check in with our tech leads. Again, this  is an editable collaborative markdown file that [17:30] lives in your enriched context space of notion.  Dan says that's fine. Just do record pointers. [17:38] Great. We can ask Claude to continue,  but consider all of the feedback from [17:43] what customers wanted from their product alongside  everything that was discussed by our tech leads. [17:50] Okay, that's a great PR, but actually I think  I'm going to have Codex give it a review. [17:55] Multi- aent orchestration. Codex takes a look  at the PR and pushes some fixes in code review. [18:05] Notion started off as a collaborative company  for humans. But at the end of the day, we believe [18:10] that we are the place for collaboration and that  collaboration is between humans with each other, [18:15] humans and agents. And as you can see in the  middle of this workflow, agents with each other. [18:20] I think if we're doing our job well, a majority  of collaboration on notion should be being done [18:24] between agents because a majority of work  in this economy is being done by agents, [18:28] especially knowledge work. And I would  argue that in order to invest in that, [18:33] you need to think about what the  right model and framework for that is. [18:37] And I would argue that having the most optionality  to do it is the best setup for your customers. [18:43] Here's what it looks like. This is I just  screenshotted our Slack. So anytime anyone has [18:47] any feedback, we drop it in this random channel  called like all feedback. We have a router that [18:54] is self-learning configured as a custom agent at  notion that will send it to the right team keep [18:59] them passively aware and then a cloud manage  agent go ahead and implement the fix. I would [19:04] posit about 40% of all of our polished task at  not all of our tasks at notion were done this way. [19:10] That is not to say that this is work that humans  were doing before. I think we're doing a lot more [19:15] work. I think our engineers are focusing on what's  the top of their license, not on things like this [19:20] like overlap between buttons. Verscell has been  really successful. They're one of our kind of [19:26] marquee customers of doing everything  from intake to launch prep to releases. [19:31] And we've seen incredible ROI from them just  within, you know, the first couple months of them [19:36] using this product. Token Town is crazy. That's  why we call it Token Town. It's the wild west. [19:43] The market is very young and I think it's very not  fair to a lot of buyers. Honestly, I think it's [19:49] quite imbalanced. I think if we all act informed  though um we can advocate for ourselves better [19:55] and actually we owe it to our customers to do  that. Our customers, our investors, each other. [20:00] Um so thank you for listening. I'm chronically  online so it's covered by the um subtitles but [20:06] serumsax at Twitter. Please DM me or reach out on  LinkedIn. I love to yap about this and think about [20:11] um how to structure problems this way. And you can  always email me. Thank you for hosting me today.