72-trial benchmark shows MCP tool output format sharply affects AI agent performance
A contributor to the CNCF Jaeger MCP server ran a structured 72-trial A/B benchmark to settle a design debate over whether MCP tools should return pre-aggregated summary rows or raw per-bucket time series data. The experiment used two production-grade AI agents — Claude Sonnet and Gemini 2.5 Pro — across six troubleshooting tasks, with three tasks designed to favor each format. Results showed that agents given time series data answered nearly all questions correctly, while those given summary rows declined to answer temporal questions seven times more often. Agents did not produce wrong answers when underfed with data — they correctly identified the limits of the information and refused to commit. The findings suggest that output format, not just token cost, is a critical and measurable factor in MCP tool design.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in