Summary
In this Latent Space podcast episode, hosts Alessio Fanelli and Swyx sit down with Simon Last (engineer) and Sarah Sachs (product) from Notion to trace the company’s multi-year journey building AI agents inside its productivity platform — from failed fine-tuning experiments in late 2022 to the recent successful launch of Custom Agents, described as Notion’s most successful product launch by free-trial conversion.
Simon and Sarah detail the technical evolution at each inflection point: designing a custom tool-calling framework before function calling existed, partnering with Anthropic and OpenAI on fine-tuning attempts that were ultimately too fragile, and finally reaching production reliability around Claude Sonnet 3.6/3.7. They walk through live use cases including automated Slack-to-task bug triaging, email-driven tenant management that enriches records with web search, and the composability patterns Notion uses internally — including a not-yet-released feature allowing agents to directly invoke other agents, slated to ship within days of recording.
The conversation covers evaluation philosophy, the practical importance of keeping humans in the loop for edge cases, MCP as a lightweight and well-permissioned integration layer, and why the highest-value enterprise agent use cases replace processes rather than people. For engineers building production agentic systems in SaaS products, this episode offers unusually candid inside detail on what it took to ship agents at Notion’s scale.
📺 Source: Latent Space · Published April 15, 2026
🏷️ Format: Podcast
