Multi-Agent System
A single agent carrying twenty tools through a long task eventually trips over itself: the context fills with half-finished threads, and the tool list alone muddies every choice. A multi-agent system divides the work between several agents, each with its own context, tools and remit.
How they are arranged
The topology is the main design decision:
- Orchestrator-worker: a lead agent decomposes the task, fans subtasks out, and assembles what comes back — the most common shape in production.
- Hierarchical: workers that are themselves orchestrators, for work that decomposes more than once.
- Handoff: control passed sideways between peers, as in the OpenAI Agents SDK, where a triage agent transfers the conversation outright.
- Shared state: agents coordinate through a common store rather than messages, the model behind LangGraph's graphs.
AutoGen, CrewAI, LangGraph and the Agents SDK each pick defaults here, which is most of what separates them in practice.
When it pays
Splitting earns its keep when subtasks are genuinely parallel — six suppliers researched at once — when one agent's output would otherwise poison another's context, or when permissions must differ so only one agent can write anything. The cost is coordination: tokens spent talking, work done twice, and a failure that is now distributed, which is why per-agent tracing stops being optional. One agent with a tight tool set beats a committee on most tasks; reach for several when parallelism or isolation is the real bottleneck.
A multi-agent system splits a task across several agents, each with its own context, tools and remit, coordinated as an orchestrator with workers, a hierarchy, peer handoffs, or through shared state. It buys parallelism, context isolation and per-agent permissions, and pays for them in coordination tokens and harder debugging. Worth it when subtasks are genuinely independent; a single agent with a tight tool set wins otherwise.
A one-person business can do everything right up until it cannot. Then you hire people, and hand each of them a slice of the work.
A multi-agent system is AI organised that way: several assistants on one job instead of a single one doing all of it. Usually a lead breaks the task up, passes the pieces out, and puts the answers back together — and each helper sees only its own piece, with only the tools that piece needs.
The gains are speed, because the pieces run side by side, and safety, because only one helper holds the keys to anything risky. The cost is all the back-and-forth of managing a team.