Passing MCP Unit Tests Doesn't Mean AI Agents Won't Break Your Server in Production

A DEV Community article argues that MCP servers, which let language models connect to tools like Prometheus, Grafana, and Loki, can pass every unit test yet still fail when driven by an AI agent. Unit tests assume a predictable caller that picks the right tool, supplies clean arguments, and calls once, while real agents may hesitate between similar tools, retry aggressively under load, and misread empty results as failures. The piece proposes a two-tier evaluation rubric, where Tier 1 checks protocol and handler correctness and Tier 2 checks behavior under an agent, with a score below 5 out of 6 on either tier signaling the tool is not production-ready. It highlights schema drift between JSON contracts and handler code as one example of a Tier 1 failure that can halt incident triage. The author adapts the framework specifically for incident response, where the gap between correct server behavior and agent behavior can be costly.