Multi-Agent System
Two or more AI agents coordinating on a task — usually through role specialization, message passing, or a coordinator agent that orchestrates the rest.
What Is a Multi-Agent System
A multi-agent system (MAS) is an architecture where two or more AI agents cooperate on a task. Each agent typically has a narrower role — a researcher, a writer, a reviewer, an executor — and they pass intermediate work between each other via messages, shared memory, or a coordinator agent that orchestrates the rest.
The pattern shows up when a single-agent loop hits a wall: context window fills with too many concerns, tool space gets too large, or the task naturally splits into parallel subtasks. Specializing roles per agent restores focus and unlocks parallelism.
Common Patterns
| Pattern | How it works | Example |
|---|---|---|
| Supervisor + workers | One coordinator delegates subtasks to specialized worker agents | OpenAI Swarm, Anthropic Subagents |
| Pipeline | Agents run in sequence, each transforming the previous one’s output | Researcher → Writer → Editor |
| Peer-to-peer | Agents debate or vote until consensus is reached | Multi-LLM debate frameworks |
| Marketplace | Agents bid on subtasks; coordinator picks the best bid | Auction-based task allocation |
When It’s Worth the Complexity
MAS is real engineering overhead. Use it when:
- The task has clear, separable subgoals (research + write + review)
- Context window in a single agent overflows past ~80% routinely
- You need parallelism — multi-agent buys you concurrent tool execution
- Tool space exceeds ~30 tools and is partitionable per role
Don’t use it when a single well-scoped agent loop with the right tools would do — most teams reach for MAS before exhausting single-agent design.
Real Implementations
- OpenSwarm — VRSEN’s open-source multi-agent runtime
- Paperclip — orchestration runtime running on homelab CT 217
- Manus — autonomous task agent with internal sub-agent decomposition
- Trinity (homelab CT 200) — referenced in the Claude MCP production review — uses supervisor + workers for long-running research tasks