A single agent with good tools can do a surprising amount. But as tasks get broader, teams reach for multi-agent systems: several agents, each with its own role, instructions and tools, coordinating toward one goal. Done well, they can tackle work that overwhelms a single agent. Done carelessly, they multiply cost and confusion.
Why use more than one agent?
- Context isolation. Each agent works in its own clean context window and returns a distilled result, instead of one window filling with everything.
- Parallelism. Independent subtasks, such as researching ten vendors, can run at once.
- Specialisation. Different prompts, tools and even models per role (a cheap fast model for triage, a stronger one for analysis).
- Separation of duties. A reviewer agent that did not write the output is a better checker than the author.
Common patterns
| Pattern | How it works | Typical use |
|---|---|---|
| Orchestrator and workers | A lead agent plans, delegates subtasks to workers, then synthesises results. | Research, analysis, broad information gathering. |
| Pipeline | Agents run in sequence; each hands its output to the next. | Document processing, content production, staged approvals. |
| Router and specialists | A triage agent sends each request to the right specialist. | Support desks, internal helpdesks. |
| Generator and critic | One agent drafts, another reviews and requests changes. | Code, contracts, reports, anything quality-sensitive. |
| Handoff | One agent transfers the conversation, with state, to another. | Customer journeys that cross domains. |
Lead agent
Plans and delegates
Workers
Run subtasks in parallel
Critic
Checks quality and policy
Result
Synthesised, verified output
The real costs
Multi-agent is not free. Each agent consumes its own tokens, and coordinating them adds more, so total usage is often several times that of a single agent on the same task. You also take on:
- Coordination failures: unclear delegation, duplicated work, conflicting assumptions.
- Error propagation: one agent's wrong output becomes another's trusted input.
- Harder debugging: behaviour emerges from interactions, so you need end-to-end tracing.
- Shared-state hazards: two agents editing the same record or file at once.
- A larger attack surface: see prompt injection, since agents pass text to each other.
A practical rule
Design tips
- Give each agent one clear job, explicit inputs and outputs, and the minimum tools it needs.
- Delegate with detailed instructions; vague hand-offs are the top cause of wasted work.
- Return concise summaries and references, not raw transcripts, between agents.
- Put budgets and stop conditions on the whole system, not just each agent.
- Evaluate the system end to end, and each agent in isolation (see agent evaluation).
Multi-agent orchestration sits inside the harness, and each agent still runs the agent loop. See our AI agent service for how we apply it.