Key takeaways
- Every retry adds latency, cost, and a chance the agent gives up.
- Most retries come from a small set of measurable causes.
- Return errors that tell the agent exactly what to change.
Retries feel harmless because they sometimes succeed. But each one adds latency, cost, and a chance the agent gives up. Cutting them is one of the highest-leverage things you can do.
Where retries come from
- Schema mismatches that fail validation the same way twice
- Ambiguous errors the agent cannot recover from
- Slow tools that time out and get called again
- Tools that return partial results the agent re-requests
How to find your worst offenders
Group calls by session and look for repeated calls to the same tool with the same arguments. Rank tools by retries per successful call. The top of that list is your fix queue.
Fixes that stick
Return errors that tell the agent exactly what to change. Accept the shapes models actually send. Cache or speed up the slow dependency behind a timeout-prone tool. Each one turns a loop into a single successful call.
See this on your own server
TrackMCP turns your MCP server's calls into adoption, workflows, and outcomes. One line to install.