Abed ManzaljiAI, Step by Step

Why a team of AI agents?

Because a lead AI that splits a big job across helpers can finish faster and better. It also costs more.

4 min to watch1 min to read

Based on a real case from Anthropic's engineering post (June 2025): one AI agent failed to list the board members of S&P 500 tech companies, and a lead AI with helpers found them.1 The screens, names and chats are made up.

Read the walkthrough as text

Without a team

  1. An analyst has a big research job, and an AI that can search the web on its own.
  2. The analyst asks for one long list.
  3. The AI says yes and gets to work.
  4. Behind the scenes, the AI makes a to-do list: one card per company.
  5. It searches the first company, Apple, and reads what comes back.
  6. Then Microsoft. Only one search runs at a time, and the rest of the list waits.
  7. Every result piles into the same notes. The longer they get, the harder it is to keep track.
  8. It's slow going. Most of the list hasn't been touched.
  9. In the end, it sends back only part of the list.

One AI, one search at a time. In Anthropic's test, a single AI agent failed this exact job, with slow searches done one after another.

With a team

  1. Same job, same question. This time the AI can call in helper AIs and act as their lead.
  2. Behind the scenes, the lead writes a plan: three slices that don't overlap, one per helper.
  3. Each helper gets a clear brief: which slice, what to bring back, and where to stop.
  4. All three helpers search at the same time, each with clean notes of its own.
  5. This is a multi-agent system: a lead AI splits the job, and helper AIs work on it side by side.
  6. The lead merges the lists, removes repeats, and checks that every name has a source.
  7. The catch: teams like this use about 15× the tokens of a normal chat, and tokens are what you pay for.

A lead AI with helpers. Anthropic's team setup beat a single agent by 90.2% on its research test, using about 15× the tokens of a chat. Worth it when a big job splits into parts.

Too big a team

  1. Early versions of this team sometimes got carried away, even on tiny questions.
  2. The analyst asks something that one search could answer.
  3. The lead AI treats it like a giant project.
  4. Fifty helpers start at once, all chasing the same single date.
  5. Every helper adds to the bill. One quick search would have done it.
  6. The fix: the lead now sizes the team to the job. A simple fact gets one AI and a few searches.

Size the team to the job. Anthropic says early versions sent 50 helpers for simple questions. Its rule now: 1 AI for a simple fact, 2 to 4 helpers for a comparison, more than 10 only for big research.

Keep going:

Why it matters, what it costs, and what to ask your team.

The real result

Better answers, bigger bill.

Anthropic built its Research feature as a team: a lead agent that plans, and helper agents that search. On its own research test, the team beat a single agent by 90.2%.1

+90.2%better than one agent on Anthropic's research test
about 15×the tokens of a normal chat
up to 90%less time on complex research questions

Team: Claude Opus 4 as lead, Claude Sonnet 4 as helpers. Compared with Claude Opus 4 working alone.

The catch: teams win partly by spending more. On one browsing test, token use alone explained 80% of the difference in scores.1

So more of me means better answers?

Often. But every helper shows up on the bill.

Try it

Worth it, or waste?

Teams shine when a job splits into parts that don't need each other. They struggle when every part depends on the rest.1,2

Sort these jobs. Is each one for a single agent, or a team?

Answers follow the guidance in Anthropic's and Cognition's posts.

Ask your team

Three questions before you build a team.

A team of agents is a cost decision as much as a technical one.

  1. Does this job split into parts that don't depend on each other?If the parts need each other's decisions, helpers produce work that clashes.
  2. Is the answer worth roughly 15× the tokens?Anthropic says teams only make sense when the task is valuable enough to pay for the extra performance.1
  3. What stops it from over-spawning?Ask for hard limits on helpers, steps and spend, and a report of what each run cost.

Under the hood, step by step: build a lead AI that splits a research job across helpers, on a budget.

Build along

Build a lead AI with helpers, one step at a time

We start with one AI working alone, run it on the same research job, and add one small thing per step. Press Run it after each step and watch the result change.

Before you ship

Five rules for a team that pays off.

From Anthropic's and Cognition's posts on building agent teams.1,2

  1. Size the team to the jobAnthropic's rule of thumb: 1 agent for a simple fact, 2 to 4 helpers for a comparison, more than 10 only for big research.1
  2. Brief like a managerObjective, output format, tools and sources, and where the task ends.1
  3. Pass references, not pilesHelpers save big results to storage and return a short pointer, so nothing gets lost in retelling.1
  4. Put the budget in codeCap helpers, steps and tokens per run. Log what each helper spent.
  5. Keep shared work togetherIf parts must agree with each other, like code, keep them with one agent.2

References

  1. "How we built our multi-agent research system" · Anthropic Engineering, Jun 13, 2025. The 90.2% result, about 4× and 15× tokens, 80% of variance from tokens, up to 90% less time, the S&P 500 board example, briefs, the 50-subagent mistake and the team-size rules.
  2. "Don't Build Multi-Agents" · Walden Yan, Cognition, Jun 12, 2025. Why split helpers make clashing decisions; the mismatched game example.

The analyst, the screens and the chats are made up; the outcomes follow Anthropic's post. The build-along output is simulated. Task sorter answers follow the guidance in both posts.

Made by Abed Manzalji, AI Engineer in Toronto.More about meContact