TLDR Tech / Practical
How AI agents behave: lessons from 63M MCP tool calls
An analysis of 63 million Model Context Protocol (MCP) tool calls made to PostHog reveals that AI agents predominantly rely on raw SQL queries and schema reading rather than typed query wrappers. Anthropic-based clients generated about half of all third-party traffic, while OpenAI's Codex exhibited the lowest error rates. Additionally, the company launched a beta MCP analytics tool to monitor robot traffic and detect errors.
Read the full story