Pattern
Verifier Loop
Every multi-agent step is checked by a deterministic verifier against the original task before the next step runs.
When to use
When the multi-agent system is producing compounding errors. When you need to bound error amplification (the DeepMind study showed verifier loops keep amplification under 2x vs 17x without). When each step's output can be rubric-graded against a known criterion.
When not to use
When the verifier itself is an LLM call with no rubric — that's not verification, it's rubber-stamping. When the task is single-step (verifier adds cost with no upside).