Multi-agent systems fail in patterns, not at random. The recent empirical literature on agent-fleet failures (Cemri et al.'s 2025 taxonomy is the reference) reads like a field guide — and like most field guides, it is most valuable before the expedition, not after.
The taxonomy groups into specification, alignment, inter-agent, and verification failures — and the inter-agent class is where multi-agent differs from single-agent engineering. Systems disobey their protocols (an agent "helpfully" edits outside its lane), echo and reinforce (reviewers agree because agreement is the conversational gravity, not because the code is good), thrash (two agents revert each other's edits in a loop), and diffuse responsibility (everyone assumed someone verified; nobody did). None of these exist in single-agent systems. All of them are protocol failures, not model failures — the members were individually fine.
| Failure mode | Signature | Structural countermeasure |
|---|---|---|
| Echo chamber | Consensus forms without new evidence | Diverse models, devil's-advocate role, judge must cite evidence |
| Thrashing | Agents revert or duplicate each other's work | Path ownership, idempotency keys, serialized writes per resource |
| Responsibility diffusion | Everyone assumed someone verified | Named owners per artifact; verification gates with audit trails |
| Protocol drift | Helpful deviations become norm | Machine-checked contracts; scope gates reject out-of-lane writes |
| Death spiral | Failed retries consume budget without new strategy | Finite retries; falsification logging; new-evidence requirement |
The field guide's meta-lesson is worth underlining: nearly every countermeasure in the right column is architectural — ownership, gates, contracts, budgets — not prompting. Fleets fail the way organizations fail, and they are fixed the way organizations are fixed: by changing the structure people work in, then the people.