Multi-agent system

A multi-agent system is a design in which several AI agents, each with its own instructions, tools and context, split a job between them and coordinate the results.

Common shapes.

• Orchestrator and workers: a lead agent plans, hands pieces to sub-agents and combines what they return.
• Handoffs: one agent passes the conversation to a specialist. In the OpenAI Agents SDK, a handoff appears to the model as a tool, so an agent named “Refund Agent” is reached through a tool called transfer_to_refund_agent.
• Fixed pipelines: code runs agents in sequence, in parallel or in a loop, as Google ADK’s SequentialAgent, ParallelAgent and LoopAgent do.

Trade-offs. Anthropic’s account of its research system (June 2025) describes the benefits: sub-agents explore in parallel, each in its own context window, and return condensed findings to the lead agent, which separates concerns and keeps any one context small. It also reports that multi-agent runs use far more tokens than a chat, and that the design fits poorly when all agents need the same context or depend heavily on each other, giving most coding tasks as an example.

Inside one system and across organizations. Most multi-agent systems run inside one application and one trust boundary, where agents can share memory and call each other in process. When the agents belong to different organizations, each stays opaque to the other. They need a protocol for discovery, tasks and authentication, which is what A2A provides. Frameworks increasingly support both: in Google ADK, a remote A2A agent can join a local agent tree as a sub-agent through RemoteA2aAgent.

Neighbouring terms. The orchestrator agent is the lead role. A sub-agent is a delegated worker. MCP connects each agent to its tools, while A2A connects agents to each other.