Multi-agent system
Several AI agents with separate roles and contexts working on one task, usually coordinated by a lead agent or a fixed workflow.
The main benefit is context isolation: each agent gets a clean, focused context window instead of one bloated conversation, and independent parts run in parallel. The costs are many times the tokens, harder debugging, and failures in the hand-offs, where one agent misreads another’s summary. A single agent with good tools is often enough; split when subtasks are truly parallel or need separate context.
Related terms
- Subagent: A separate agent run with its own fresh context, started by a main agent for one subtask.
- Orchestrator-workers pattern: One model plans the task, hands parts of it to worker models or subagents, and combines their results.
- A2A (Agent2Agent protocol): An open protocol for one AI agent to hand a task to another agent, possibly from another vendor, and follow it until it is done.
- Tokens: The pieces of text models read and write, and the unit you pay for.
- Agent control plane: The layer that starts, watches, limits and stops many AI agents at once: queues, permissions, budgets, logs and a view of who is doing what.
