← AI Terminology
Multi-Agent System
A multi-agent system is an architecture where multiple AI agents — each with its own context, tools, and goals — collaborate, communicate, and coordinate to accomplish tasks too complex for a single agent, with different agents specialising in different aspects of a problem.
It is the dominant paradigm for complex agentic AI workflows in 2024–2025.
It is the dominant paradigm for complex agentic AI workflows in 2024–2025.
Why It Matters in AI
A single LLM agent has limited context, can only do one thing at a time, and can't specialise. Multi-agent systems overcome this: a coder agent, a tester agent, a reviewer agent, and an orchestrator can work in parallel, each with focused context, and cross-check each other's work. They enable longer, more complex tasks than any single context window allows and introduce redundancy and specialisation that improves output quality — but also introduce coordination failures and error propagation risks.
Key Points
| Aspect | Description |
|---|---|
| Roles | Orchestrator (planner), specialists (coder, researcher, writer), critic (reviewer) |
| Trust | Orchestrators must not blindly trust sub-agent outputs — verification and sandboxing matter |
| Frameworks | AutoGen (Microsoft), CrewAI, LangGraph, Swarm (OpenAI), AgentTorch |
| Parallelism | Independent sub-tasks can run simultaneously — speedup proportional to task decomposability |
| Communication | Message passing between agents — shared memory, tool outputs, direct prompts |
| Failure modes | Error propagation, agent disagreement, infinite loops, coordination overhead |
Simple Analogy
A law firm on a complex case: the partner (orchestrator) assigns research to a junior (research agent), drafting to another (writer agent), and fact-checking to a third (critic agent) — each working in parallel, reporting back, with the partner synthesising and directing. No single person does everything; coordination is the key challenge.
Common Usage Examples
- AutoGen:
AssistantAgent + UserProxyAgent— Microsoft's conversational multi-agent framework - CrewAI:
Crew(agents=[researcher, writer, editor], tasks=[research_task, write_task, edit_task]) - LangGraph: stateful graph where nodes are agents and edges define routing logic
- Claude multi-agent: sub-agents with
computer_usetool perform parallel web research swarm(OpenAI): lightweight agent handoff framework —client.run(agent=triageAgent, messages=[...])
Summary
In short: Multi-agent systems coordinate multiple specialised AI agents to tackle tasks too complex for a single agent — enabling parallelism, specialisation, and longer-horizon work, at the cost of coordination complexity.