Member-only story
I did the obvious thing. I connected a stack of MCP servers to a single agent — filesystem, GitHub, Slack, a browser, a search tool, a database client, a calendar — and expected a Swiss-army-knife superpower. Instead the agent got measurably worse. In a controlled tool-selection stress test, a plain LLM picking the right tool out of a large Model Context Protocol pool scored 13.62%. Not 90%. Not 50%. Thirteen percent. And before it did a single useful thing, it had already burned tens of thousands of tokens just reading the tool menu.
That number — 13.62% — is not a typo, and it is not mine alone. It comes from the RAG-MCP paper (arXiv:2505.03275), and once I saw it I could reproduce the shape of the curve on my own machine. The uncomfortable finding: MCP tool count is not a free upgrade. Past roughly 20 tools, agent reliability doesn’t gently taper. It falls off a cliff.
Here’s why more tools make your agent dumber, the exact numbers behind the collapse, and the two fixes — one retrieval-based, one Anthropic’s code-execution pattern — that clawed accuracy and tokens back.
Why this is suddenly everyone’s problem #
MCP won. It’s the default way to give an agent hands, and the ecosystem exploded — every SaaS has a server now, and the natural instinct is to bolt on all of them “just in case.” That instinct is exactly the trap.