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

PostHog’s analysis of 63 million MCP tool calls from 130,000 people over 90 days found that AI agents, primarily operating through code editors and terminals rather than chat interfaces, are increasingly interacting with product data and features. After excluding PostHog’s own automated traffic, Anthropic-based clients accounted for roughly half of third-party calls, followed by OpenAI at 25%, with Claude Code and Codex leading individual clients. Agents relied heavily on raw SQL and schema inspection while making limited use of the platform’s many specialized tools, though they created substantially more analytics insights, dashboards, actions, and experiments than human web users. Codex-generated requests had the lowest error rate among third-party clients, while SQL queries caused more failures than schema discovery, and agents often successfully recovered by retrying failed calls. The study also found that a small number of dashboard-related operations generated a disproportionately large share of returned tokens. In response, PostHog introduced MCP Analytics, a beta monitoring feature that records tool-call events, client activity, errors, response sizes, and agent sessions to help MCP providers identify usability problems and improve agent-facing tools.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.