Multi-agent system
A multi-agent system uses multiple LLM-driven agents working together — typically with role specialisation (planner, executor, critic) or domain specialisation (researcher, coder, writer). Coordination patterns include hierarchical (orchestrator delegates), peer-to-peer (agents communicate as equals), and pipeline (each agent processes and passes to the next).
Multi-agent systems trade complexity for capability. The wins: specialisation lets each agent be optimised for its task (better system prompts, narrower tool sets, smaller context); parallelism lets independent sub-tasks run concurrently; modularity lets one agent be improved without disrupting the others. The costs: orchestration overhead (the coordination loop has its own LLM cost), debugging difficulty (failures span multiple agents), and the temptation to over-decompose (every step becomes a separate agent when one would suffice). Healthy multi-agent design starts with the smallest viable agent count and adds specialisation only where it earns its keep.
Related terms
- Orchestrator agent
An orchestrator agent is the top-level controller in a multi-agent system — it decomposes the user's request into sub-tasks, dispatches them to specialist agents, integrates their outputs, and handles cross-cutting concerns like error recovery and progress reporting.
- Specialist agent
A specialist agent in a multi-agent system is scoped to a narrow domain — code review, data analysis, document drafting, web research — with a system prompt, tool set, and context window tuned to that domain.
- Autonomous agent
An autonomous agent is an AI system that pursues a goal over multiple steps with minimal human intervention — deciding what actions to take, executing them via tools, observing results, and iterating until the goal is achieved or judged unreachable.