Notion’s Token Town: 5 Rebuilds, 100+ Tools, MCP vs CLIs and the Software Factory Future — Simon Last & Sarah Sachs of Notion

Notion rebuilt its agent system five times between 2022 and 2024, iterating through non-existent tooling, poor model reasoning, and short contexts before reliable function calling and improved models made custom agents viable. The team emphasizes a portfolio approach to product development—balancing maintenance, near-term wins, and speculative AGI-pilled projects like software factories where agents collaboratively build, debug, and maintain codebases. They frame Notion not as a wrapper but as a collaboration-layer platform akin to Datadog’s relationship to AWS: leveraging foundational infrastructure while delivering differentiated UX expertise.

It reveals how a leading product-led company navigates the gap between early AI promise and shippable reality, with hard-won lessons on when to persist, pivot, or wait for model improvements.

1. 01Reliable AI agents required both better models (e.g., GPT-4 with function calling) and deep product iteration on permissions, context scoping, and background execution.
2. 02Notion’s 'portfolio approach' to R&D balances shipped product refinement, working on imminent capabilities, and investing in long-term bets like autonomous software factories.
3. 03The company sees its role as the system of record for collaborative work—not building hardware or agents from scratch, but integrating and enriching data from external tools.