Multi-agent system

OpenAI describes two broad patterns. In the manager pattern, a central model calls specialist agents as tools and combines what they return. In the decentralised pattern, agents hand execution to one another². The OpenAI Agents SDK makes the same split. A manager can keep the conversation and call specialists as tools, or a handoff can make the specialist the active agent for the rest of the turn³. Anthropic describes an orchestrator-workers workflow. A central model breaks the task into subtasks that are not fixed in advance, delegates them to worker models and synthesises the results¹.

This is a step beyond one model calling several tools. OpenAI says extra agents can separate roles when instructions get too complicated or tools get confused with one another, and that different tasks may suit different models². The run still needs an exit. In the guide, a loop continues until a final output, an error or a maximum number of turns, and a multi-agent run can include a series of tool calls and handoffs before that exit².

Anthropic also describes parallelisation, where independent subtasks run at the same time, or the same task is run more than once and the outputs are aggregated¹. Both labs say not to add agents by default. Anthropic says to add complexity only when it improves the outcome¹. OpenAI says more agents bring complexity and overhead, so a single agent with tools is often enough, and to split only when one agent fails to follow complicated instructions or keeps picking the wrong tool².

A 1995 survey by Wooldridge and Jennings treats interaction with other agents, through a communication language, as one part of a weak notion of agency. The paper presents itself as a survey of work already under way⁴.