A multi-agent system is an AI application in which two or more agents work on one task. Each agent has its own instructions, tools and context, and the agents hand work and results to each other, usually through a coordinating agent. The goal is to split a problem that is too broad, too long or too parallel for a single agent.
Each participant is an AI agent in the ordinary sense: a model, tools, a loop and a stop condition. The multi-agent part is how they are connected.
How a Multi-Agent System Works
Three architectures cover most systems in production:
- Orchestrator and workers. A lead agent plans the task, starts worker agents for the subtasks, and combines what they return. The workers do not talk to each other. Anthropic’s research feature works this way: a lead agent plans, spawns subagents that search in parallel, and decides whether more research is needed (Anthropic engineering).
- Handoff. One agent finishes its part and transfers the conversation to a specialist, for example from a triage agent to a billing agent. Only one agent is active at a time.
- Pipeline. Agents run in a fixed order, each with a narrow role: draft, check, format. This is closer to a workflow than to free collaboration.
Anthropic’s guide to agent design lists orchestrator-workers among its core patterns and recommends adding such complexity only when it measurably improves results (Building effective agents).
Agents built by different vendors or teams can also be connected over a network protocol such as the Agent2Agent Protocol, which lets one agent discover another and delegate a task to it without sharing its internals.
Why Split the Work
- Parallelism. Independent subtasks, such as researching ten companies, run at the same time.
- Clean context. Each worker sees only what its subtask needs, so its context window is not filled with other agents’ tool output.
- Specialisation. A worker can have a short prompt, a small tool set and a cheaper model that suits its job.
What It Costs
Multi-agent systems spend much more on tokens. Anthropic reports that in its data agents use about four times the tokens of a chat interaction, and multi-agent systems about fifteen times. The same article reports that its multi-agent research setup outperformed a single agent by 90.2% on its internal research evaluation. Both figures come from one vendor’s workload; they show the shape of the trade-off, not a rule.
When Not to Use One
The same source lists cases where several agents fit poorly: tasks where every agent needs the same full context, tasks with many dependencies between steps, and most coding tasks, which have fewer independent parts than research. Low-value tasks rarely justify the extra cost. If a single agent with good tools solves the problem, it is easier to test, cheaper to run and simpler to debug.
Common Failure Modes
- Lost information at handoffs. A worker returns a summary that drops the detail the lead agent needed.
- Duplicate work. Two workers get overlapping instructions and search the same sources.
- Error amplification. A wrong claim from one worker is treated as fact by the next.
- Hard debugging. A bad final answer can start anywhere in the chain. Without a trace of every agent’s steps, nobody can find where.
In Practice
In our AI agent development work, a second agent is added only when an evaluation shows that splitting the work beats one agent with better tools. When the split pays off, the lead agent gives each worker an explicit objective, output format and boundary, and workers return structured results rather than prose. The control flow lives in an orchestration layer, usually LangGraph (see agent orchestration), and one LangSmith trace follows a request across all agents, so a reviewer can find which agent introduced an error.