multi-agent-patterns

Place a central supervisor agent in control that decomposes user objectives into subtasks and routes them to specialist workers. Implement a forward_message tool so sub-agents can pass responses directly to users, avoiding the telephone game problem where supervisor paraphrasing loses fidelity.

Supervisor architectures centralize control in one orchestrating agent, giving strict workflow control but creating a bottleneck. Swarm architectures let agents hand off directly to each other via transfer functions, removing single points of failure but requiring explicit convergence constraints.

Multi-agent systems consume roughly 15 times the tokens of a single-agent chat baseline, compared to about 4 times for a single agent with tools. Token usage explains about 80 percent of performance variance, so the added capacity often justifies the cost for complex tasks.

Avoid multi-agent designs for simple queries or tasks that do not decompose into parallel subtasks, since coordination overhead and token costs negate benefits. Upgrading to a stronger single model often provides larger gains than adding agents.

Use weighted voting that favors higher-confidence or domain-expert agents instead of simple majority voting. Debate protocols with adversarial critique rounds and sycophancy triggers that detect agents mimicking each other also reduce false consensus.