PostHog's 38 lessons: build for agents before you build an agent

PostHog has built 10 AI products in four years. Its MCP server, the interface that lets outside AI agents use PostHog, is used daily by 5x more people than its own assistant, PostHog AI. Its AI install wizard does about two hours of setup in roughly 8 minutes, and wizard users convert to paid 5x more often. The team says context, not the agent's harness, is the moat. Why it matters: It's a rare cost and usage readout from a team shipping agents at scale. Expose everything users can do as agent tools and skills, give agents room to write their own SQL and code, and trace every model call from day one. Reliability beats new capabilities, and most agent failures turn out to be missing context. One caveat is that the 5x conversion figure is observational, and people who choose the wizard may already be more motivated.

Try this: List your product's actions that agents can't do yet, and ship the missing tool or skill for the most-requested one before building any assistant UI.