When Anthropic first built an AI that could do research on its own, they gave it a team.
One model, the lead, would read the question, break it into parts and send a handful of helper models off to search the web in parallel. Then it would gather what they found and write the answer. On paper, it was a newsroom.
In practice, it was a newsroom on a new editor’s first day. Early versions spun up 50 helpers for simple questions. Some helpers searched endlessly for sources that didn’t exist. Others kept interrupting each other with updates.
The briefs were the real problem. The lead would send out an instruction like “research the semiconductor shortage.” One helper went off to study the 2021 car chip crisis. Two others both researched 2025 supply chains, running the same searches side by side. Nobody had told them who was covering what.
The brief is the job
A lead AI splits the job among helper AIs. How clearly it briefs them shapes the result.
Anthropic named it “orchestrator-workers” in a 2024 guide: a central model breaks a task down on the fly, hands the pieces to worker models and stitches together what comes back. The point is that nobody decides the pieces in advance. The orchestrator works them out from the task in front of it.
When it works, it works well. Anthropic’s finished research system, with Claude Opus 4 as the lead and Claude Sonnet 4 models as helpers, beat a single Claude Opus 4 by 90.2% on the company’s internal research tests. Asked to find the board members of every information technology company in the S&P 500, the team split the list and found them. The single model, searching one company after another, did not.
An editor on election night works this way. She doesn’t write every story. She gives one reporter the governor’s race, another the turnout numbers, a third the ballot measures, and spends her own time deciding what goes on the front page. Her skill is in the splitting and the briefing.
One of Anthropic’s main fixes, alongside work on tools and evaluation, was to teach the lead to brief like that editor. Every helper now gets four things: a goal, the shape the answer should come back in, which tools and sources to use, and where its task ends. The lead also got rules of thumb for how big a team to hire: one agent for a simple fact, a crowd only for a sprawling question.
Where it breaks
The pattern has a price and a shape it fits.
The price is volume. In Anthropic’s data, a multi-agent run used about 15 times the tokens of an ordinary chat, so it only makes sense when the answer is worth that much.
The shape is work that splits cleanly. Anthropic warns that tasks where every agent needs the same context, or where the pieces depend heavily on each other, are a poor fit, and that most coding work has fewer truly parallel parts than research does. A Google study in December 2025 found the same split. Teams of agents did well on a financial analysis task that broke into independent parts. On a planning task where each step depended on the last, every multi-agent setup they tested did worse than a single agent. It is Brooks’s Law again: more hands help most when the work can be divided.
The manager you already are
Picture the Monday you hand out a market report in four pieces, one line each. By Friday, two colleagues have both pulled apart the same competitor’s annual report, nobody has touched pricing, and the fourth is polishing a section you never needed. Nobody slacked off. The brief was one line long.
Anthropic’s four-part brief is a good test for any handoff, human or AI. Does each person know the goal, what to hand back, what to use, and where their part stops? When one of those is missing, the duplication and the gaps tend to follow.
The AI version adds two twists. In Anthropic’s system the lead couldn’t steer its helpers once they set off, so whatever was missing from the brief stayed missing. And the bill for an oversized team arrives immediately, in tokens. The early system that sent 50 helpers after a simple question was making a manager’s oldest mistake at machine speed: hiring before deciding what the work was.