A taxonomy of how agent teams actually break in production, drawn from running a real pipeline (Herald → Atlas → Gauntlet → Forge → Cipher) across a dozen projects. These aren’t hypotheticals. Each one has cost me a session.
1. Context drift
Each agent paraphrases the previous handoff in its own words. After four hops, “build a login form with email + password, defer OAuth” becomes “implement full auth including OAuth providers.” Nobody lied. Each step was a 5% restatement. The compounding is the bug.
Mitigation: structured handoff blocks that are copied, not summarized. The original task brief stays pinned at the top of every agent’s context.
2. Mock divergence
The QA agent (Gauntlet in my setup) writes tests against a mocked dependency. The build agent (Forge) later changes the real interface. Tests still pass because they’re testing the mock. Ship, break, page.
Mitigation: contract tests that hit the real interface at least once per merge, plus a rule that mocks live next to the code they mock so a rename breaks compilation.
3. Unowned escalations
An agent discovers the task is larger than scoped. It writes “this should probably be a separate feature” in its output and proceeds anyway, or punts entirely, assuming the parent will catch it. The parent sees a completion signal and moves on. The escalation goes into the void.
Mitigation: escalations are a first-class output type, not a comment. If an agent writes “scope change” anywhere, the orchestrator pauses and routes to a human.
4. Silent partial completion
The agent finishes 80% of the task, hits a snag, wraps up with “implemented the core logic” and declares done. The other 20% — the annoying 20% — is now invisible technical debt.
Mitigation: verification hooks that check every acceptance criterion from the task brief, not just “did the file change.”
5. Over-cautious escalation
The opposite failure. An agent could easily handle a decision, but escalates every small choice to the human because refusing to decide feels safer than deciding wrong. Death by a thousand “should I use tabs or spaces.”
Mitigation: explicit authority in the agent prompt. “You own formatting decisions. You do not own architecture decisions.” Draw the line in writing.
6. Hallucinated dependencies
Agent imports a package that doesn’t exist, or calls an API method the library never had. The model pattern-matched to something plausible. Tests catch this — if you run them. Which is why verification hooks aren’t optional.
The pattern across all six: agent failure modes are coordination failures, not intelligence failures. The models are good enough. The scaffolding around them is what decides whether the team ships.