$building.effective.agents
Menu
Last verified: April 2026
· Multi-agent systems

Multi-agent systems.

A multi-agent system is a system in which two or more LLM-based agents collaborate, usually under a coordinator. The orchestrator-worker pattern is the simplest case; more elaborate architectures introduce role specialisation, message buses, and shared memory.

Why multi-agent

Anthropic's “Building Effective Agents” treats multi-agent as a special case of orchestrator-worker, with the recommendation: use a multi-agent system only when the task structure justifies it. Most production cases that look like “multi-agent” can be handled by a single agent with a larger toolset and a clearer plan. The cases that genuinely benefit from multi-agent are those in which:

  • Distinct competences (researcher, planner, executor, critic) warrant distinct prompts and possibly distinct models.
  • The agents must operate concurrently with shared state, not just be invoked in sequence.
  • The system models a problem that is itself multi-actor (a negotiation, a debate, a multi-party simulation).
coordinatortask graphresearchersearches + readsplannerbreaks into stepsexecutorcalls toolsshared busagents publish + subscribe; the coordinator owns scheduling

Coordination patterns

Hierarchical (orchestrator-worker)

One coordinator decomposes work and delegates to specialised workers. This is the orchestrator-worker pattern, named by Anthropic and documented in detail. The coordinator owns scheduling, the workers own execution. Most production multi-agent systems are this shape.

Conversational (peer-to-peer)

Two or more agents exchange messages. The seminal academic reference is CAMEL (Li et al., 2023), which introduced communicative agents that role-play to solve a task collaboratively. The framework analogue is AutoGen (Wu et al., 2023), which models multi-agent conversation as the primitive.

Simulation (many-agent)

Larger populations of agents acting in a shared environment. Park et al., “Generative Agents” (2023) simulates a small town of LLM-driven agents with memories, plans, and reflection. The simulation use case is well-documented; the production-system use case rarely scales to many agents because coordination overhead grows quickly.

Communication models

Multi-agent systems differ in how agents exchange information. Three patterns appear in framework documentation:

  • Shared memory. Agents read from and write to a common state object. Simple to reason about; coordination overhead is low. LangGraph models this as a typed state on the graph.
  • Direct messaging. Agents send messages to each other. AutoGen models this as a conversation history.
  • Publish-subscribe bus.Agents publish events to a bus; other agents subscribe to the events they care about. CrewAI's flow events use this shape.

Public examples

  • CAMEL and the underlying paper: reference for role-playing dual-agent systems.
  • AutoGen: multi-agent conversation framework from Microsoft Research.
  • CrewAI: role-and-task multi-agent framework.
  • LangGraph multi-agent docs: graph-based multi-agent shapes (supervisor, network, hierarchy).
  • OpenAI Swarm: experimental routines and handoffs for multi-agent.

Cost considerations

Multi-agent cost is the sum of per-agent costs plus the coordinator's own LLM call. Compared to a single-agent equivalent, multi-agent typically costs more for two reasons: duplicated context (each agent receives some shared context anew) and coordination overhead (the coordinator's call is pure overhead). Vendor docs on prompt caching reduce the duplication cost; nothing reduces the coordination cost without reducing the agent count.

Failure modes specific to multi-agent

Multi-agent systems inherit single-agent failures and add their own:

  • Coordination overhead exceeds task value. The coordinator spends more compute managing the agents than the agents save by parallelising.
  • Disagreement deadlocks. Two agents reach contradictory conclusions and the coordinator cannot pick. A tie-break rule is required up-front.
  • Context fragmentation. Each agent has only part of the picture; decisions are locally optimal but globally wrong.

See failure modes for the broader taxonomy and orchestrator-worker for the canonical multi-agent pattern.

Glossary

See multi-agent system, coordinator, worker, message bus.

Read next