How AI agents actually use your MCP server: five failure modes you won't see in logs
The author instrumented a demo ticket-booking MCP server with a self-built tool called mcpspan and identified five kinds of agent invocation problems that never show up in logs: agents guessing tool names that don't exist; parameters that don't match the schema and get rejected by the SDK before the handler runs; parameter types misunderstood because of how the tool description is worded; retry loops that keep hitting the same parameter; and responses so large they eat into the context window.
Why it matters: The author used a self-built MCP analysis tool to empirically surface five failure modes in agent calls that logs don't reveal, and these can be adapted to troubleshoot your own MCP server.
