A free guide by Pierre Sabbagh
Run 100 agents at once in Claude Code
One keyword, ultracode, makes Claude Code write a team of subagents for your task. How to ask for 100, watch them, and keep the run inside your plan.
Needs a computer and a paid Claude plan. You won’t write any code.
It picks one job worth a big run and writes what to paste. Press send when it opens.
Not now? Send it to my computer.
Written November 17 from Claude Code’s own docs. The steps are below.
Let AI walk you through it
It reads this guide and works through it with you, one step at a time.
You’ll need a paid Claude plan for Claude Code. Dynamic workflows are on all paid plans.
When your browser asks, allow it to open the app. The prompt will be typed in for you, so just press send.
Need the app first? Get the Claude app for Claude Code.
Or follow the guide yourself
Here’s how to do what I showed in my video. In Claude Code, you put the word ultracode in a prompt, and Claude writes a small program that runs many subagents for you. You can ask for 100. You won’t write any code.
What you need
- A paid Claude plan. This feature, called dynamic workflows, is on all paid plans. On Pro, turn it on first: type
/configand switch on the Dynamic workflows row. - Claude Code, not the regular chat, kept up to date.
- A folder of real work to point it at, such as a project with many files.
Step 1: ask for a workflow
Type ultracode: and then your task, in Claude Code:
ultracode: audit every API endpoint under src/routes/ for missing auth checks
Claude Code highlights the keyword, and Claude writes a workflow for the task instead of working through it one step at a time. The keyword only changes how Claude organizes the work. The agents get the same permission checks as anything else in your session.
Asking in your own words (“use a workflow”, “run a workflow”) also works.
If you didn’t mean to start one, press Option+W on a Mac or Alt+W on Windows to dismiss the highlight for that prompt.
Step 2: say how big the team is
A subagent is a helper that takes one piece of the job. Claude aims for a team size set in /config under Dynamic workflow size:
| Setting | Claude aims for |
|---|---|
| small | fewer than 5 agents |
| medium | fewer than 10 agents |
| large | fewer than 50 agents |
| unrestricted | whatever fits the task |
The default is medium, or small if you’re on Pro. This is advice, not a limit. A prompt that asks for a different size overrides it, so you can write:
ultracode: spin up 100 subagents and review every file in this repo for bugs, then give me one report
Step 3: approve the plan
Before it runs, Claude Code shows the plan with its phases. In the terminal, choose Yes, run it, View raw script to read what Claude wrote, or No. In the desktop app you get a card with Once, Always and Deny.
In auto mode, you’re asked only on the first launch, and not at all if you’ve turned on /effort ultracode for the session. In bypass permissions mode, it starts without asking.
Step 4: let it run, and watch
It runs in the background, so your session stays usable. Type /workflows to see each phase, how many agents are in it, and each agent’s token use. Press p there to pause the run, or x to stop it.
When it finishes, the result comes back into your session as one report.
How big a run can get
- Up to 1,000 agents in one run. 100 is well inside that.
- 16 run at the same time by default, fewer if your computer has fewer CPUs. The rest wait their turn. You can change that with the
CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTSsetting, from 1 to 256, on Claude Code v2.1.269 or later. - You can’t type into a run halfway. It pauses by itself only for an agent’s permission prompt or when you hit a usage limit. If you want a sign-off between stages, run each stage as its own workflow.
Keep it inside your plan
A big run uses a lot of your plan. Runs count toward your usage limits.
- Start on a small slice, one folder instead of the whole repo, and watch the token numbers in
/workflows. You can stop the run there, usually without losing the work already done. - Look out for the
Large workflowwarning. It shows when a run goes past 25 agents or about 1.5 million tokens. It only warns: it doesn’t stop anything. - Check
/modelfirst if you normally switch to a smaller model for routine work. Agents use your session’s model unless told otherwise. You can ask Claude to use a smaller model for the easy stages. - If you hit your usage limit, the run pauses and carries on by itself after the reset, as long as the app stays open. This needs a recent version (Claude Code v2.1.271, or v2.1.319 in the desktop app’s Code tab), a subscription without usage-based billing, and a limit that resets within 24 hours. It waits at most twice. After that, agents fail, and you can ask Claude to relaunch the workflow.
The other way: ultracode for the whole session
The keyword runs one task. To have Claude plan a workflow for each substantive task in a session, type /effort ultracode. Turn it off with /effort ultracode off. Each request then costs more tokens and time, and you’ll reach a session or weekly limit sooner. It also skips the large-run warning and, in auto mode, the first-launch approval.
Which jobs fit
The docs’ examples are jobs with many similar pieces: checking every file for the same problem, migrating hundreds of files, and research where sources get cross-checked against each other. For a task that is one small change, a normal prompt is faster and cheaper.
Written on 17 November 2026 from Claude Code’s own page on dynamic workflows (code.claude.com/docs/en/workflows) and its model configuration page. I haven’t run these steps myself while writing this guide. Claude changes fast: if a menu has moved, look for the same words in the app.
The prompt
This is what the buttons send. Copy it and paste it into the AI app you use.
Walk me through this guide, one step at a time: https://pierresabbagh.com/resources/run-100-agents-in-claude-code Read the whole page first. If your web tool can't open it, try a command like curl. If you still can't open it, tell me so I can paste the text in, and don't guess the steps. Explain each step in plain words, ask before you install or change anything, and stop after each step so I can check it. The guide is written for Claude Code. If you are a different app, tell me how to do each step in your app, and say so if you're not sure it works there.
Send this guide to your inbox.
Next guide
Make your Claude Code limit last with the Advisor
Set Sonnet as your main model and let it call Opus only at the hard moments, with Claude Code’s Advisor.
At work
Doing this at a company?
With a team, the hard part is the task itself, the people who do it and who checks the result.