Multi-agent system

• Tasks require distinct expertise (data cleanup, drafting, critique).
• You want resilience—one agent can recover when another fails.
• You need exploration and debate to improve answer quality.

Multi-agent systems can improve quality but add latency, cost, and complexity. PMs choose roles, communication protocol, and arbitration rules. They must measure marginal value of additional agents versus user patience and budget.

Start with two roles (doer + critic) and a clear arbiter. Cap turns and enforce a budget. In 2026, run bandit-style routing: skip critics when confidence is high, engage them on risky intents. Log disagreements to improve prompts and tools.

A design review bot uses a builder and a critic agent. On complex UI tickets, quality score improves 12 points with only 350 ms added latency; on simple tickets the critic is skipped automatically, saving cost.

• Letting agents debate without turn limits, causing slow or stuck sessions.
• Not measuring incremental value per agent, leading to cost bloat.
• Sharing unfiltered context between agents, risking data leakage.