How I run six Claude Code agents at once
Friday hits. I close the laptop. The agents don't stop.
While everyone logs off, my screen has a bunch of tabs open. Each one with an agent working alone, on a different task. I stop. They keep going.
Looks like magic. It's not. It's architecture. And I'm going to show you the real thing, bugs included, not the pretty screenshot.
First: what an agent actually is
An agent is not a chat. That's the confusion that traps 90% of people.
The chat gives you an answer and dies right there. You ask, it answers, done. The agent is a different animal. It's a loop with three parts:
- Tools. What it can touch (read a file, run a command, open the browser, make a git commit).
- Context. What it knows about the task and about your project.
- Stop criteria. When it considers the work done and hands it back to you.
The chat answers a question. The agent chases a goal until it hits the stop criteria. It reads, tries, fails, reads the error, tries again. On its own.
Once you separate that in your head, you stop treating AI like a question box and start treating it like a process. And a process you can run in parallel. A question, you can't.
How I cut the work into pieces
The trick is not making the AI code faster. It's in how you split the task before you let the agents loose.
Each agent gets a closed scope. No shared state with the others. It returns a structured result that I stitch together at the end. Something like this:
agent 1 → writes the tests for the payment module
agent 2 → refactors the CSV parser
agent 3 → hunts the timezone bug in the report
agent 4 → documents the public API
Notice: none of them depends on the others. That's on purpose.
It's the same principle as a good team. You don't put two people editing the same file at the same time, because it turns into merge conflicts and fights. You cut the problem into independent pieces and let each one run alone in its own lane.
The rule I follow: if two tasks touch the same file, they do not run in parallel. Either I merge them into one, or I run them in sequence. Shared state between agents is the number one source of silent screw-ups.
The workflow in practice (fan-out and barrier)
Here's how it actually works. I use a workflow that does a fan-out of the independent tasks: it fires them all at once, each agent in its own tab. The ones that don't depend on each other run in parallel. The ones that do depend wait at a barrier until the previous result lands.
┌─ agent 1 ─┐
fan-out → ├─ agent 2 ─┤ → barrier → I review and merge
├─ agent 3 ─┤
└─ agent 4 ─┘
The practical result: on Friday afternoon I close the laptop and four, five, six fronts keep moving. Monday I show up, read what each one delivered, and keep what's good.
The part nobody tells you
Now the brutal honesty, because the rest is course salesmen.
Running six agents in parallel is not six times faster. I used to think it was. I got burned.
There's the agent that gets stuck in a loop and keeps hammering the same error until you kill the tab. There's the agent that "finishes" the task with the test red and swears it's done. There's the agent that makes a dumb decision in the first 30 seconds, and the next 10 minutes are built on top of a wrong foundation. This happens. To me, every week.
That's why the bottleneck stops being typing and becomes reviewing. The more agents you let loose, the more output lands on your desk to check. If you don't review, you've just multiplied the speed at which you produce bad code. AI doesn't take you off the hook. It pushes you up the chain: you stop being the one pressing the key and become the one deciding what's good.
Two things that saved me a lot of pain:
- Small, closed scope per agent. Task too big = lost agent = rework. Break it down more than you think you need to.
- Explicit stop criteria. "Done when the tests pass," not "done when you think it's good." A machine has no shame about lying.
It's not about typing faster. It's about no longer being the execution bottleneck and becoming the person who architects and reviews. That's the real game.
Want me to show you the fan-out workflow on screen, agent by agent? Drop a comment below.
follow me for more real stuff from someone who uses AI for actual work, not for posting pretty screenshots.