Your User Isn't Human: Testing MCP Servers

When Your User Isn't Human: Testing MCP Servers

All tests are green and coverage is high — yet the agent tells the user something like: "I don't have a tool for that." The tool exists. It's covered by tests. And the bug isn't in any single line of code. 

That's the big surprise MCP servers hold for QA: their primary user is not a human or another service, but an AI agent. The model doesn't "connect to your API" — it reads. Everything it sees about your server is text: the tool's name, its description, and its parameter schema. That means the contract now consists not only of the schema but also of the text — and changing a single word in a description changes the system's behavior in production. 

In this talk, we'll break down how an agent chooses a tool and where the scenario falls apart: the agent didn't pick your tool, called it with the wrong arguments, or misunderstood the response.