Multi-agent system

For a research task, a coordinating agent might assign one agent to gather sources and another to check claims, then assemble their results. In a peer design, those agents exchange findings directly and negotiate who handles a gap. These are two possible coordination patterns, not requirements to use different underlying models.

Multi-agent systems are broader than language-model products. Agents in an auction can compete, while robots can coordinate movements toward a shared goal. What matters is the interaction between decision makers. Several model calls in a fixed sequence do not automatically establish separate agents.

Splitting work can make responsibilities clearer or allow independent tasks to run concurrently. It also creates handoffs where information can be lost, work duplicated or disagreements left unresolved. Define what each agent owns, what messages mean and who decides when the task is complete. Compare the arrangement with a single-agent baseline before accepting its extra cost.