A multi-agent system is an arrangement in which several AI agents, each with its own role, instructions, and tools, work together toward a shared goal. Rather than one agent handling an entire task, the work is divided across specialists that coordinate — by passing control, reporting to a supervisor, or sharing state — so that the combined result is more focused or more capable than a single agent could achieve alone.
Why multi-agent systems matter
As a task grows, a single agent strains against two limits. Its instructions become a tangle of competing concerns, and its context fills with detail relevant to only part of the work, which degrades the model’s focus and raises cost. Asking one agent to research, write, fact-check, and format a report means every turn carries all four jobs at once.
A multi-agent system addresses this by decomposition. Each agent gets a narrow role, a tailored prompt, and only the tools it needs, so each stays focused and easier to evaluate. The trade-off is coordination: agents must exchange information, control has to pass cleanly between them, and the overall run now has more moving parts that can fail or stall. Whether the decomposition is worth it depends on whether the task genuinely has separable sub-problems — added agents that do not correspond to real boundaries usually add overhead without adding capability.
How it works
Multi-agent systems differ mainly in how coordination is structured. Common arrangements include:
- Supervisor (or orchestrator) — a coordinating agent decomposes the task, delegates sub-tasks to specialists, and integrates or reviews their results.
- Sequential pipeline — agents run in a fixed order, each consuming the previous agent’s output, suited to tasks with clear stages.
- Swarm — peer agents hand off control to one another as the task evolves, with no central coordinator deciding every step.
- Router — an initial classifier inspects the input and dispatches it to the agent best suited to handle it.
Across these patterns, two mechanisms recur: a handoff, which transfers control and context from one agent to another, and shared state, which lets agents read and contribute to a common record of progress.
Single-agent vs. multi-agent
The choice is one of decomposition, not capability. A single agent keeps everything in one context and one set of instructions, which is simpler to build, trace, and reason about. A multi-agent system distributes the task, gaining focus and modularity but introducing coordination cost and more places to fail. A practical rule is to start with one agent and introduce more only when distinct sub-problems, conflicting instructions, or an overloaded context make the boundaries real.
In practice
Coordinating several agents reliably depends on the same durability and visibility a single agent needs, applied across the whole group. A durable, observable runtime persists each agent’s steps and the handoffs between them, so the system resumes after a crash and every transfer of control is recorded. This builds on the AI agent as the unit, uses handoffs to pass control, and relies on orchestration to coordinate the parts. For the available arrangements, see multi-agent strategies.