An agent that can query production telemetry can test its own hypotheses against reality, rather than hallucinate fictions. Observability tooling allows our agents to close the loop and benefit from production wisdom. Today, more than 40% of Honeycomb's weekly active users run production queries through our MCP server via their agents.

But an MCP server must do more than wrap existing APIs to be legible to the agents. Getting token economy, tool descriptions, or schemas wrong makes access technically present but practically useless. I'll walk through the output format choices, decisions from evals, and bugs (solved and still being found) that made ours trustworthy in production, enabling users and their agents to see clearly in production.