From Note-Taking App to AI Workspace: The Simon Last Interview

Summary

• Notion’s AI strategy has advanced from writing tools to autonomous work. GPT-4 access in 2022 led to AI Writer in February 2023 after 2 or 3 months; workspace Q&A reached GA in October 2023; and the personal agent shipped last fall, around August or September, after Notion had tried to build an agent three or four times. The investor signal is iteration speed rather than a single model bet: Notion rewrites its AI harness “probably every 6 months or so.”
• Retrieval quality—not workspace neatness—is the craft layer beneath useful enterprise agents. Embeddings make hierarchy less important because the AI only needs to find the right snippet, but chunk sizing, pipeline design, and source-specific tuning remain hard. In response to Sarah’s observation that native products have not solved this perfectly, Simon says most companies are “pretty bad at making their indexes”: Slack and Google Drive cannot share a one-size-fits-all retrieval strategy.
• Coding agents are widening talent dispersion far more than shrinking team size. Notion still prefers small tiger teams, but Simon argues a tool-fluent developer can now be “a 100 or 1,000x engineer”: “the minimum bar has not changed, but the maximum bar has increased dramatically.” The catch is operational—agents can produce ambitious, thoroughly tested work or “all slop,” depending on architecture, verification, review, and safe deployment.
• Notion is repositioning from software humans operate to a workspace where humans manage and authorize autonomous agents. Its personal agent inherits a user’s access, while custom agents begin with none and can be authorized to monitor Slack, file tasks, research, or work with reports in the background. Simon’s larger thesis is that code lets agents bootstrap missing capabilities: “I think of coding agents as the kernel of AGI.”
• The platform strategy is deliberately model-neutral as model quality and economics keep shifting. Simon calls Notion “the Switzerland for models,” arguing customers should be able to switch as different labs lead in different months. He says four Chinese open-source models were already “quite good” and “way cheaper” than frontier models; Notion had added one the prior week and planned to offer all four.
• The lived workflow is already managerial, with autonomy earned through an approval-to-trust loop. Simon runs Claude Code or Codex agents overnight, has kept one coding agent working for 13 days, and says, “I haven’t written code since last summer.” His email agent moved from interviewed preferences and daily corrections to automatically archiving the roughly 95% of messages he does not need—the clearest example of becoming “the agent manager instead of the coder.”

Deep dive

Not yet available upstream; scheduled sync will retry.