The most dangerous MCP bug I hit was not dramatic. No outage siren. No heroic war room. Just one renamed field and a workflow that started making slightly worse decisions with a straight face.
That is why I am bullish on MCP as a protocol and deeply suspicious of MCP rollouts without contract tests. The protocol gives agents a cleaner way to use tools. It does not magically make your tool contracts stable. If anything, it makes contract drift easier to spread because now more agents can depend on the same broken assumption.
The failure
A tool response changed from a flat owner field to a nested owner object. Reasonable backend cleanup. Unfortunately, one agent path still expected the old shape and silently fell back to free-text inference.
That fallback worked in demos. It failed in the places that matter: ambiguous ownership, handoff workflows, and anything involving three teams plus a Slack thread that looked like a hostage note.
The practical impact in our sample runs:
- owner-routing accuracy dropped from 92% to 76%
- manual review rose by roughly 14 points
- the average failure took longer to debug because every individual trace looked "plausible"
Plausible is the worst kind of wrong. Obviously broken systems get fixed. Plausibly wrong systems become quarterly OKRs.
What to enforce in CI
- request/response schema validation for every tool action
- capability negotiation checks for optional features
- backward-compatibility snapshots for high-use payloads
- negative fixtures for missing fields, unknown enums, and null payloads
Integration Strategy: Ad Hoc vs Contract-Tested MCP
The conservative MCP take
MCP is a great direction because it standardizes the boring boundary. But standard boundaries are not self-enforcing. You still need versioning, compatibility policy, and tests.
My rule now: if a tool is important enough to expose through MCP, it is important enough to have contract tests before an agent touches it.
Takeaway
MCP will make agent tooling more portable. It will not save you from sloppy contracts. Treat MCP servers like production APIs, not like a clever plugin folder, and the rollout gets much less spicy.
