Key takeaways
- Agents chain calls, so a slow tail compounds across a session.
- Track p50 for the typical call and p95 for the tail.
- Fix latency at the tool level, not the server average.
Latency matters for agents differently than for humans. An agent chains calls, so a slow tail on one tool compounds across a session and pushes it toward timeouts and retries.
Why averages lie
A mean latency looks reassuring while a tenth of calls take ten times as long. Those tail calls are the ones agents time out on. Track percentiles, not averages.
The two numbers to watch
- p50: the experience of a typical call
- p95: the tail that quietly breaks sessions
Act at the tool level
Sort tools by p95, not overall latency. The fix is usually one dependency behind one tool. Speeding up that tail often lifts completion rate more than any broad performance work.
See this on your own server
TrackMCP turns your MCP server's calls into adoption, workflows, and outcomes. One line to install.