Skip to content
SaytPro
Say hello

Guide

Agent teams: five Claudes in one room

This is the system we promised in our Instagram video. A single agent's answer sounds logical, but you do not know whether it is right. In a Claude Code agent team, 5 roles debate: one is the judge, while the others defend theories and try to prove each other wrong with your numbers. You get the answer left standing.

Updated
Reading time4 min read
You needClaude Code, a paid Claude plan
01

Install Claude Code

claude.ai

One command in the terminal, or the Claude desktop app. You need a paid Claude plan.

02

Turn on agent teams

settings.json

You add a few lines to the settings file. The feature is still experimental and off by default.

03

Put the numbers in one folder

CSV, Excel

Export your reports as CSV or Excel files. The team cites only the numbers in these files.

04

Paste the prompt

Claude Code

The roles try to disprove each other's theories with numbers. The answer left standing comes to you as a report.

01 · Claude Code

Install Claude Code

Claude Code is a program from Anthropic that lets Claude work with the files on your computer. You install it with one command in the terminal. If the terminal is new to you, you can also use the Claude desktop app.

Once it is installed, you type claude in the terminal and sign in to your Claude account in the browser. Claude Code needs a Pro, Max, Team, Enterprise or Console account. The free claude.ai plan does not include Claude Code.

macOS, Linux

curl -fsSL https://claude.ai/install.sh | bash

Windows (PowerShell)

irm https://claude.ai/install.ps1 | iex

02 · settings.json

Turn on agent teams

Agent teams are still an experimental feature in Claude Code and are off by default. To turn them on, add these lines to your settings file: ~/.claude/settings.json on macOS and Linux, .claude\settings.json in your user folder on Windows. If the file does not exist, create it.

If the file already has other settings, do not delete them: add the env section next to the settings already there.

settings.json

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Claude Code docs: agent teams (opens in a new window)

03 · CSV, Excel

Get the numbers ready

The team works only with the numbers you give it. Create a new folder and put the reports that relate to your question in it: your list of leads, a report from your ad account, an export from Google Analytics or Search Console. CSV and Excel files are the easiest to work with.

If the files contain customers' names, phone numbers or email addresses, delete those columns first. Dates, sources and counts are enough for the analysis.

Then go to this folder in the terminal and type claude.

04 · Prompt

Paste the prompt

Paste the prompt below into Claude Code. Replace the two parts in square brackets with your question and the names of your files. You can adapt the roles to your own field, but keep the judge: it defends no theory and only checks the numbers.

Prompt: agent team

Create an agent team. Question: [For example: why did leads from our website go up last month?]
Data: the files in this folder [for example: leads.csv, ads.csv, analytics.csv].

Create five teammates and give them these names:
1. marketer: looks for the cause in advertising and campaigns.
2. designer: looks for the cause in the site's design and ease of use.
3. skeptic: checks whether this is chance, seasonality or a measurement error.
4. psychologist: looks for the cause in how customers decide.
5. mathematician: defends no theory and acts as the judge. Checks every claim only against the numbers in the files.

Rules:
- Every theory must cite at least one number and the file it comes from.
- In round 1, everyone writes their own theory. In round 2, they message each other and try to disprove each other's theories with numbers, as in a scientific debate.
- The mathematician marks every theory the numbers disprove as OUT and gives the reason in one sentence.
- Do not make up any number that is not in the files. If there is not enough data, say so plainly.
- Teammates must not change the files, only read them.

At the end, give me a short report: the theory left standing and the numbers that support it, the theories that are out and why, and the next step to check this.

While the team works

Claude becomes the team lead and creates the teammates. They appear in a panel below the input box: select one with the arrow keys and press Enter to see its work and message it directly.

If Claude uses ordinary helper agents (subagents) instead of a team, ask again and say clearly that you want an agent team.

When the debate is over, the team lead writes the report. Double-check the numbers behind the theory left standing against the files yourself: the report is a starting point for a decision, not the decision itself.

Careful. A team uses far more tokens than a single chat, because each teammate is a separate Claude session. For a simple question, a single chat is enough. Save the team for questions where the cause is unclear and there could be several explanations.

Checklist

WhatWhat to check
QuestionOne question you can check with numbers, such as "Why did leads go up?" Not a general question like "How is my business doing?"
DataThe reports cover the same period, with the previous period included for comparison.
Personal dataCustomers' names, phone numbers and email addresses have been removed from the files.
JudgeOne role defends no theory and only checks the numbers.
ReportYou find the numbers behind the theory left standing in the files and check them yourself.

Your site's numbers

Does your site track leads?

The team is only as good as the numbers you give it. If your site does not track leads and where they come from, there will be nothing to debate.

On the sites SaytPro builds, a new lead goes straight to the admin panel and to a Telegram channel as a notification.

Where do you see the leads that come from your site?

Frequently asked questions

The guide and the prompt are free. Claude Code itself needs a paid Claude plan (Pro, Max, Team, Enterprise) or a Console account. The free claude.ai plan does not include Claude Code. Because a team uses more tokens than a single chat, you reach your plan's limit sooner.

No. Installing takes one command, and the prompt is written in plain language. You do not write any code: Claude reads your files and gives you the result as text.

The Claude Code docs recommend 3 to 5 teammates for most tasks. Of the 5 roles in the prompt, one is the judge and the rest defend different theories: the debate stays rich and the cost stays manageable.

A single agent usually stops looking once it finds one convincing explanation. In a team, each teammate defends its own theory and tries to prove the others wrong, so the answer left standing is more likely to be right.

The files stay on your computer. Claude reads them through your account and sends their contents to Anthropic to produce the answer. So remove customers' personal data from the files beforehand.

Check that the lines in the settings file are written correctly and restart Claude Code. If Claude still uses ordinary helper agents, say clearly in the prompt that you want an agent team.