Multi-Agent Systems
Multiple AI agents working together, each handling a specialised role in a larger workflow.
A multi-agent system is an architecture where two or more AI agents collaborate — each with its own role, tools, and context — to complete tasks too complex or large for a single agent.
Common patterns:
- Orchestrator + Subagents: One coordinator agent breaks a task into subtasks and dispatches them to specialised agents. Used in Claude Code: the orchestrator plans, subagents write files, run tests, and report back.
- Pipeline: Agent A's output feeds Agent B's input in a fixed sequence (research → outline → draft → edit)
- Peer collaboration: Agents debate or critique each other's outputs — reduces hallucination, improves reasoning quality
Why multi-agent over one agent?
- Context limits: A 1M-token task split across 4 agents, each with 250K context, is more tractable
- Parallelism: Independent subtasks can run simultaneously
- Specialisation: A coding agent + a testing agent + a documentation agent outperform one generalist trying all three
Engineering challenges: Agent communication format, handling failures mid-pipeline, preventing infinite loops, cost management across many parallel agents, and security (one compromised subagent can corrupt the whole system via prompt injection).
In plain terms
A project team vs. one person doing everything. The PM coordinates, the designer designs, the developer builds, the QA tests — each expert in their lane, communicating through defined handoffs.