Contents
At a glance
- Category
- Core agent concept
- Common topologies
- Supervisor and specialists, sequential pipeline, peer collaboration, critic
- Real benefit
- Modularity - shorter prompts, smaller tool lists, testable boundaries
- Not a benefit
- Raw reasoning quality; more agents do not think better
- Cost
- One model call per handoff, with latency stacking sequentially
- Required control
- A global step budget across all agents, plus explicit shared state
What a multi-agent system is
A multi-agent system divides a job across several agents, each with its own instructions, tools and often its own model. A coordinator - variously called an orchestrator, supervisor or router - decides which agent handles what and passes results between them. The alternative is one agent with all the tools and one long system prompt.
The honest framing is that multi-agent is a modularity decision, not a capability upgrade. Two agents do not reason better than one; they are easier to reason about. Each has a shorter prompt, a smaller tool list and a testable boundary, which is the same reason you split a large function into small ones.
The common topologies
Supervisor is the workhorse: a router reads the input, picks a specialist, and returns the result. Inbound enquiry goes to the sales agent, complaint goes to the support agent, invoice question goes to the accounts agent. It is easy to test because each path is independent, and easy to extend because adding a specialist does not touch the others.
Sequential pipeline passes work along a chain - research, then draft, then check - which suits content and review workflows and is really a agentic workflow with agent-shaped steps. Peer collaboration, where agents debate or critique each other, produces better quality on genuinely hard reasoning and multiplies cost; it is rarely the right answer for a business process. A separate critic agent that checks one output against explicit criteria is the version of that idea that does earn its cost.
Cost and the failure modes
Every handoff is another model call with its own context, so a three-agent pipeline costs roughly three agents' worth plus the orchestrator. Latency stacks the same way, sequentially. That is fine for a nightly batch and a problem for anything a customer is waiting on - see latency budget.
The characteristic failures are handoff loops, where two agents pass a task back and forth because neither considers it theirs, and context loss, where agent B never learns the crucial detail agent A saw. Both are solved the same way: an explicit shared state object rather than passing prose between agents, plus a global step budget across the whole system rather than per agent. Without a global cap a loop between two capped agents is still unbounded.
What you should know
Start with one agent. Split only when you can name the reason: the tool list has grown past what one model chooses from reliably, two parts of the job need different models or different permissions, or two teams need to own different pieces. Splitting for elegance costs money and produces a system where nobody can say which agent caused a bad output.
A2A is worth knowing about here - it is a protocol for agents from different vendors to discover and delegate to each other, which is a different problem from orchestrating agents you built yourself. For frameworks, compare LangGraph and CrewAI, which take opposite positions on how much structure to impose.
Multi-agent system FAQ
What is a multi-agent system?
Several specialised agents, each with its own instructions and tools, coordinated so they can hand work to each other. A supervisor or router usually decides which agent handles a given input and assembles the result.
Is a multi-agent system better than one agent?
Not inherently. It is more modular, which makes each part easier to write and test, but it costs a model call per handoff and introduces coordination failures a single agent cannot have. Split when one agent's tool list or prompt has grown unmanageable, not for elegance.
What is the most useful multi-agent pattern for a business?
Supervisor and specialists. A cheap router classifies the incoming item and dispatches to the right specialist agent. Each path is independently testable, adding a new specialist does not risk the existing ones, and the router itself can run on a small, fast model.
Why do my agents keep passing work back and forth?
Because their responsibilities overlap or neither has a clear rule for accepting the task. Fix the boundaries in the prompts, make the router the only component that assigns work, and add a global step budget across the whole system - per-agent caps do not stop a loop between two agents.
How do agents share information?
Through an explicit shared state object rather than by passing prose. If agent B only receives agent A's summary, it will eventually lose the one field that mattered. Write structured results to shared state, and give each agent a defined read and write surface on it.
What is the difference between a multi-agent system and A2A?
A multi-agent system is an architecture for agents you build and orchestrate yourself. A2A is an open protocol for agents built by different parties to advertise capabilities and delegate to each other across organisational boundaries. One is internal design, the other is interoperability.
Primary source
LangGraph - agent architecture concepts
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what multi-agent system means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
