What AI agents actually do with your MCP server

Description and use

MCP servers may appear healthy, yet agents encounter hidden failures such as missing tools, bad arguments, misleading descriptions, looping calls, and large responses that inflate token usage.

Problem

MCP servers appear healthy but agents encounter hidden failures such as missing tools, bad arguments, misleading descriptions, looping calls, and large responses that inflate token usage.

Outcome

By measuring agent interactions, the team uncovered hidden failure modes that were not visible in standard logs, leading to concrete fixes that improved agent reliability and reduced token costs.

Workflow

• Instrument server with mcpspan
• Run simulated traffic
• Collect tool call data
• Analyze missing tool calls
• Identify argument validation failures
• Adjust tool descriptions
• Detect and fix looping calls
• Measure response sizes
• Implement limits or summaries

Evidence

Agents call tools you don't have

Source

Bad arguments are the most common failure, and they hide

Source

The tool description is code

Source

Agents loop, and every single call looks fine

Source

Response size is invisible on latency charts

Source