WTF is Multi-agent System?
Several LLM agents, often with different roles or tools, working together on one task.
Also called: mas, multi-agent, agent swarm
TL;DR
Instead of one agent doing everything, you have a researcher, a coder and a reviewer, or many parallel searchers. Splitting the work can improve quality and speed, but adds cost and new ways to fail (agents talking past each other).
How it works
Topologies include orchestrator–workers (hierarchical), peer-to-peer conversation, and pipelines. Benefits come mostly from parallelism and separate context windows per agent, which let the system cover more ground than one context could hold. Costs: token usage multiplies, errors can compound across agents, and debugging is harder. Interoperability protocols like A2A aim to let agents from different vendors cooperate.
Origin
- Coined / introduced by
- Classic AI concept, revived for LLMs
- When
- 2023
- Story
- Multi-agent systems are decades old; LLM versions took off in 2023 with role-playing frameworks such as CAMEL.