Multi-agent systems
Several agents with separate contexts and roles cooperate, usually a lead agent delegating to sub-agents.
Changes fast: check the latest updates too. Reviewed 17 Oct 2026.
Key ideas
- Best fit: broad, parallelisable work like research across many sources.
- Sub-agents protect the lead agent's context by returning condensed results.
- Costs multiply quickly; multi-agent research runs can use many times the tokens of a single agent.
- Give each sub-agent a precise brief: objective, output format, tools, and boundaries.
- Poor fit: tightly coupled tasks where agents must share lots of state, such as most coding edits.
Common mistakes
- Splitting into many agents to look sophisticated when one agent with good tools would do.