Single agents face hard limits: context window size, sequential execution speed, and the generalist/specialist tradeoff. Multi-agent systems address these by distributing work across agents that operate in parallel, specialize in their subtask, and each operate over a smaller, more manageable context.
Understands the overall goal. Decomposes it into sub-tasks. Assigns tasks to subagents. Monitors progress. Integrates results into a final output. Knows the big picture.
Receive a scoped sub-task. Complete it using their tools and return the result. No awareness of the broader goal. Can be highly specialized — a code review agent, a research agent, a summarization agent.
Central message bus — Agents publish and subscribe to events. All inter-agent communication is logged and auditable.
Shared memory store — A database all agents can read and write. Simpler but requires conflict management when multiple agents write concurrently.
Linear pipeline — Agent A's output becomes Agent B's input. Simple and debuggable for tasks with a natural sequential structure.
Multi-agent systems are more complex to build, debug, and operate. Errors propagate. Coordination adds latency. Parallel agents multiply API costs. They are justified when tasks genuinely benefit from parallelism, require deep specialization, or exceed a single agent's context window capacity.
Ask the AI assistant about multi-agent architectures, the orchestrator-subagent pattern, or how to manage shared state across agents.