Multi-agent systems split a hard problem across several specialized models working as a team.
ConceptWhat it is
Multi-agent systems assign different LLM agents distinct roles, such as researcher, coder, and critic, that communicate to jointly complete a task no single generic prompt handles well.
They exist because decomposing a complex goal into specialized sub-agents, each with a focused prompt and toolset, often produces better and more inspectable results than forcing one agent to juggle every responsibility at once.
How it worksThe mechanics
A coordinator or shared message thread routes a task to specialized agents in sequence or in parallel; each agent operates within its own context and toolset, produces output, and passes it to the next agent or back to the coordinator, which decides when the overall goal is complete.
At a glanceSee it
The design decision behind multi-agent — fan work out only when subtasks are genuinely independent and the quality gain outruns the added token and latency cost.
The orchestrator-worker mechanism — a planner decomposes the goal, specialist agents run in parallel with their own tools, and a synthesizer merges their partial results into one answer.
When to use itWhere it fits
- Complex tasks that naturally decompose into specialized roles.
- Workflows benefiting from a built-in critic or reviewer agent for quality control.
- Long, multi-stage content or code-generation pipelines.
- Situations where parallelizing independent sub-tasks across agents saves time.
When NOT to use itLimits & anti-patterns
- Simple tasks solvable by one well-prompted agent, where extra agents add cost without benefit.
- Latency-critical applications, since inter-agent handoffs multiply round-trip time.
- Systems lacking clear coordination logic, which can lead to agents talking past each other or looping indefinitely.
Trade-offsAdvantages & costs
Advantages
- Specialization often improves quality on complex, multi-faceted tasks.
- Built-in critic roles catch errors before they reach the user.
- Easier to reason about and improve one narrow agent role at a time.
- Supports parallelism for independent sub-tasks.
Trade-offs & costs
- Coordination overhead and inter-agent communication add cost and latency.
- Debugging emergent multi-agent behavior is harder than a single agent.
- Risk of agents reinforcing each other's errors without proper checks.
- Significantly more moving parts to design, test, and monitor.
ExampleIn the real world
An automated software-engineering platform uses a multi-agent setup where a planner agent scopes the task, a coding agent writes the implementation, and a reviewer agent checks the diff against tests before merging.
ToolsHow to implement it
- AutoGenMicrosoft framework for building conversable multi-agent systems.
- CrewAIrole-based framework for orchestrating agent teams.
- LangGraphmodels multi-agent handoffs as an explicit graph with shared state.
- OpenAI Swarmlightweight pattern for agent handoff and coordination.
Cost & effortWhat it takes
Cost multiplies with the number of agents and handoffs, each consuming its own tokens. High engineering effort to design coordination, guardrails, and failure handling.
What changedWhat changed here
Updated this page Multi-agent collaborations can collapse when agents with opposed objectives lack a shared goal, so governance must be designed in.
On sub/agt-multi-agent-systems, add that agents need a shared goal or governance because opposed objectives without one make the interaction collapse.
- Claude Code relaunches Projects to manage multiple AI agents in the cloud
Claude Code relaunched Projects, letting you run multiple agents under one roof with shared memory, goals, and a file/artifact library, coordinated by a 'coordinator' with parallel threads. That's a concrete shift in how you'd structure multi-agent coding work — shared context instead of isolated sessions — and worth testing before you build your own orchestration layer.
- Claude Code Gets Major Update: AI Sessions Can Now Communicate Autonomously, Marking a Key Step in Multi-Agent Collaboration
Claude Code now lets AI sessions communicate autonomously, a concrete step toward multi-agent collaboration for builders.
Three kinds of claim, strongest first. Signal runs every morning.