Definition of the multi-agent system
“A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together.” The sentence comes from an Anthropic article of 13 June 2025. The documentation of the LangChain framework speaks of systems that coordinate specialised components (texts consulted on 12 November 2026).
The expression is older in computing: according to the website of the IFAAMAS foundation (ifaamas.org), the AAMAS conference, devoted to autonomous agents and multi-agent systems, was born in 2002 from the merger of three conferences. This entry sticks to the recent sense: several AI agents based on a language model, each of which receives its own instructions, tools and context.
Multi-agent system, single tool-equipped agent, fixed chain of calls: not to be confused
A single tool-equipped agent remains a single agent, however many tools it has. In a fixed chain of calls, the code sets the order of the steps and the model handles each one. Anthropic’s article “Building effective agents” (19 December 2024) calls these systems with predefined code paths workflows, and reserves the word agents for those in which the model directs its own process. The documentation of OpenAI’s Agents SDK likewise contrasts orchestration by the model with orchestration by code, which it describes as more deterministic and more predictable.
Orchestrator, specialised agents, handoff
Two forms of cooperation recur in the documentation consulted.
- The orchestrator and its specialised agents. A main agent breaks down the task, distributes it, then assembles the results. LangChain calls this pattern subagents: the main agent calls the others as tools, and all routing passes through it. CrewAI offers a so-called hierarchical process, in which a manager agent assigns tasks and validates results.
- The handoff. One agent passes the task on to another, which continues it. In OpenAI’s Agents SDK, this handoff is presented to the model as a tool; AutoGen’s documentation describes the same pattern.
The MCP protocol connects an AI application to tools and data; Agent2Agent (A2A) presents itself, on its website (a2a-protocol.org, consulted on the same day), as an open standard for communication between AI agents.
What splitting the work brings, and what it costs
LangChain’s documentation puts forward three reasons for splitting the work: conserving the context window, letting separate teams each maintain their own agents, running subtasks in parallel. The price can be read on three levels.
- Tokens. Anthropic’s June 2025 article reports that, in the vendor’s data, multi-agent systems used about fifteen times more tokens than a chat: a figure specific to that vendor, which does not amount to a rule.
- Coordination. The same article describes agents that, lacking precise delegation instructions, duplicated the same work or left gaps.
- Verification. Two runs of the same request may follow different paths: evaluation then focuses on the result. The trace must follow the task from one agent to the next: this tracking is a matter for LLMOps.
One risk is added: what one agent passes on enters the context of the next, so that an error, or an instruction slipped into content read upstream, can spread.
An orchestrator and three agents for a freight quote
Imagine a freight forwarder that draws up quotes for international transport. An orchestrator agent receives the request and breaks it down. A “rates” agent queries the carriers’ rate grids. A “customs” agent consults, read-only, the regulatory documentation applicable to the goods. A “drafting” agent assembles the quote, which an employee validates before it is sent.
If every request followed the same steps, a fixed chain would suffice. Splitting the work is justified here by requests whose course changes with the goods, the route and the documents supplied.
Our AI agents page explains how to choose between an orchestrated workflow, a tool-equipped agent and a multi-agent system, then how to set permissions and human supervision.