Loading...
How multiple agents reconcile disagreements (debate, chairperson moderation), the characteristic failure modes of agentic systems (cascading tool failures, role drift, infinite handoff), and the orchestration patterns that mitigate them.
Multi-agent systems amplify both the power and the failure modes of single agents. A debate of three agents converges on better answers than a solo agent — when it converges. When it doesn't, you get unbounded loops, role drift, and a frustrating user-facing experience. This chapter is the practical how-to: how to coordinate multi-agent teams, how to detect and recover from cascading failures, and which orchestration patterns are robust enough for production.
This chapter covers:
Click any topic to jump in
N agents critique each other across rounds. 5–15pp accuracy gain on problems with knowledge but weak commitment.
Explicit moderator role — picks speakers, declares consensus, terminates the discussion.
Spec drift, tool misuse, hallucinated tools, infinite loops, premature termination — and the multi-agent extras.
How one bad observation poisons N downstream steps. Circuit breakers contain the blast radius.
Step budgets, deadlines, fallback chains, idempotency keys — the production hardening that makes coordination safe.
Multi-agent debate (Du et al., 2023; Liang et al., 2023) runs agents in parallel on the same question, has them critique each other across rounds, and aggregates to a final answer. The technique improves factual accuracy and calibration, often by 5-15 percentage points, on problems where the model has the knowledge but commits to wrong answers under direct prompting.
A typical debate has three rounds:
After rounds, the answer with the most agent votes wins (or a separate judge model picks). Empirically, rounds with agents is the sweet spot. Diminishing returns kick in at .
Why it works: each agent's first answer is a confident guess. The cross-examination forces them to defend — and good defenders update under good critiques. The aggregate is more robust than any individual.
Debate works best when the agents have different priors:
Three identical agents with identical prompts produce three identical answers and do not improve under debate. Variance is a feature.