A multi-agent system divides a larger job among several software agents. One may retrieve facts, another plan the work, another perform a task and another check the result. Messages or a shared work queue pass information between them, sometimes under the direction of an orchestrator.
This arrangement can help when tasks require different tools or can run independently. It does not guarantee improvement over one well-designed agent. More agents usually mean additional model calls, longer traces, greater coordination overhead and more opportunities to lose context.
Errors may propagate between agents. An early inaccurate summary can become the next agent's assumed fact, and a malicious instruction hidden in retrieved content can be repeated through the system. Tool permissions and the provenance of messages matter at each boundary.
Start with a single-agent baseline, then add roles only where testing shows a benefit. Measure the final task outcome, cost, latency and recovery from failures, and keep logs that let reviewers reconstruct the handoffs.