How AI agents behave: lessons from 63M MCP tool calls

PostHog studied 63 million AI agent calls to its tools, found most came from code editors, and built a way to replay them.

PostHog makes product analytics software and offers its features to AI agents through a standard connection protocol. Natalia Amorim reports on 63 million tool calls from 130,000 people over 90 days. Most calls came from agents working inside code editors and terminals rather than a chat window.

• Only about 7 percent of the calls came from a chat window.
• Running a database query made up a quarter of all calls, and reading what data exists made up another 7 percent.
• Dashboard tools drew 0.5 percent of calls but used 18 percent of the text the AI models processed.
• PostHog built a replay of each agent session, call by call, like its replays of human sessions.

A product team can treat AI agents as a user group to observe, with their own analytics and replays. This is a vendor's account of its own product, and it describes a new monitoring feature rather than a design change.

Derived by Working Surface from the article; more in the Playbook. Source line: If your newest power user is a robot, you should probably know what it's doing.

2 October 2026: teams building with AI agents are deciding in advance, and writing down, where the agent stops and a person checks: in product code, in what the product promises users, and in visibility of agent behaviour.