Multi-agent systems

1 · What it is

Here, an agent is a language model that works in a loop. It picks a next move. It calls a tool, such as web search, to act or look something up. Then it reads what came back and chooses again. A multi-agent system runs several of these loops and links them up. OpenAI’s Agents SDK, a toolkit for building agents, sets each one up with three things: its instructions, the tools it may call, and handoffs, which let it pass work to another agent.

There are a few common wirings. A manager can treat other agents like tools. It gets one piece back from a specialist, then writes the final reply itself. A handoff works the other way round. The specialist takes over and replies for the rest of that turn. Or a plain program can set the order. It can feed one agent’s output into the next. It can also run several side by side on tasks that do not depend on each other. Setting the order in code makes run time and cost easier to predict.

Research papers have tried other shapes. In a CAMEL demo, one chat agent plays a stock trader and keeps giving instructions to another that plays a programmer, and together they work on a trading bot. AutoGen split the work between two agents. An assistant writes code. A “user proxy” stands in for the person: it runs that code or asks a real human for input. In multiagent debate, several copies of one model each answer a question. Each reads the others’ answers and revises its own, and they usually end up agreeing. The authors found better maths and strategy reasoning, and fewer invented facts, known as hallucinations. A Berkeley study later sorted the ways these systems fail into three groups. They were system design, agents falling out of step, and checking whether the task was really done.

So when is one agent better? Think of a group project at school. Splitting it up helps when each person can research a different chapter on their own. It hurts when everyone needs the same notes or waits on each other. Agent teams work the same way. Models read and write text in tokens, small chunks of text. A single agent already uses about four times the tokens of a normal chat. A team uses about fifteen times, so the job has to be worth that bill. Agents are also not yet good at handing work to each other on the fly. Teams earn their cost on valuable jobs that split into many parallel parts, need more reading than one agent’s context window (the text it can hold in view at once) can take, or use many complicated tools.