You already know how to manage AI agents. You've been doing it with people for years. Someone put it to me that way on a call last month, and it's the most useful thing anyone's said to me about AI this year.
If you've ever run a small team, hired a freelancer or handed a job to a colleague, you've already solved most of the problem people think needs technical skill.
The short version
- Two or more AI tools on the same work go wrong for the same reasons two people do: unclear jobs, no shared list, bad handovers and two pairs of hands on one document.
- The fix is management, not code. Give each tool one job, keep one list of who's doing what, hand over in writing and never let two of them edit the same file at once.
- AI agents need these habits more than people do, because they don't remember yesterday and they don't ask when they're stuck.
- You can start in under a minute with one line per tool.
Why do two AI tools on the same job get messy?
Here's how it usually goes. You've got ChatGPT drafting something, Claude tidying it up and maybe Gemini or Copilot pulling in research. Each one is good on its own.
Put them on the same piece of work and it falls apart. One rewrites what the other just fixed. Nobody knows which version is current. You end up copying text between chats, holding the whole job in your head and wondering why three clever tools feel slower than doing it yourself.
That's not a tool problem. It's a team with no manager. The tools have changed. They now run longer jobs on their own, in your files, without you watching every step. That's what moved the problem from "write a better prompt" to "run the team".
If that framing sounds familiar, it's the difference between doing the work with AI and designing what AI does, which I wrote about in Operator or Orchestrator: Two Very Different Ways to Use AI.
The four habits you already use with people
1. Give each one a job
You wouldn't hire two people and tell them both to "help out". You'd say who does the books and who answers the phone.
Do the same with your AI tools. One drafts, one checks. One researches, one writes. In my own business, Claude Code and Codex both work from the same set of business files, and each one checks the other's work. Neither of them is there to "help out".
A reviewer that knows it's the reviewer behaves differently from one asked for its opinion. I went into why in How to Make AI Catch Problems in Your Work.
2. Keep one list of who's doing what
Every team that works has one place where the jobs live. A whiteboard, a shared spreadsheet, a rota on the fridge.
AI tools need the same thing, written down where all of them can read it. Mine is a single file listing what's live, what's waiting and who owns the next move. Every AI that works in my business reads it before it starts.
Without it, each tool starts from its own guess about what matters. With it, they're working from one list, not three opinions.
3. Hand over in writing
When someone goes on holiday, a good colleague leaves a note: what's done, what's half-done, where the files are, what to watch out for.
AI tools need that note every single time, because the one picking up the job doesn't know what happened yesterday unless someone wrote it down. In my setup, when one AI finishes a piece of work it leaves a short note in an inbox folder saying what it did and what it read. The next one starts from that note instead of from scratch.
If you've already turned a repeat job into written instructions, you've done the hard part. That's the same muscle I described in The Work You Repeat Every Week Should Be a Skill.
4. Never two on the same file at once
Two people editing the same document at the same time is how you lose an afternoon's work. Everyone's been there with a shared spreadsheet.
AI tools do it faster and without noticing. So set a rule: one tool changes a file at a time, and the others suggest changes rather than making them. My shared memory file goes further. It has a lock, so if one AI is writing to it, the next one waits its turn.
You don't need a lock to start. You need the rule, written where every tool will read it.
What changes when the team is AI
The habits carry over. Three things about the team don't.
They don't remember the way people do. A colleague remembers Tuesday's conversation on Wednesday. An AI agent usually starts each job fresh, and even the chat tools with a memory feature decide for themselves what to keep. Anything that matters needs writing down where they'll read it. That's why the list and the handover note matter more here than they do with people.
They don't ask. A new starter who's stuck will usually come and find you. An AI that's stuck tends to guess and carry on, and it sounds just as confident when it's guessing.
They're fast. A person who misreads the brief wastes a morning. An AI that misreads it can redo half your files before you've finished your coffee. Clear jobs and one-at-a-time editing stop a small mistake becoming a big one.
None of that needs technical skill. Every rota, job sheet and handover note you've ever written carries straight over. What changes is how strictly you stick to it.
Where to start this week
Open a blank document and write one line for each AI tool you use on the same work:
- What its job is.
- What it's not allowed to touch.
For example: "ChatGPT writes first drafts only. It never edits the final version." Short and plain is fine.
That's it. Then paste the line into the place each tool reads before it starts: Custom instructions in ChatGPT's settings, Instructions for Claude in Claude's settings, or Your instructions for Gemini under Personal context. For an agent like Codex or Claude Code, it goes in the instructions file it reads, AGENTS.md or CLAUDE.md.
Add one more phrase to the end of each line: "Do this every time, even when I ask for something quick." AI tools tend to let a standing rule slip when a short, specific request comes in, and that's exactly when two of them end up on the same work.
Next time two of them are on the same job, you'll see the difference straight away. They stop treading on each other because you've told them where to stand.
Common questions
Do I need technical skills to run several AI tools together?
No. The hard part is deciding who does what and writing it down, which is a management job. The technical setup can come later, once you know what you want each tool to do.
What if two AI tools give me different answers?
Treat it like two colleagues disagreeing. Look at where they differ, because that's usually where the real decision sits. Neither tool gets the final say. You do.
How do I stop two AI tools changing the same document?
Give one of them the job of making changes and tell the others to suggest edits instead. Write that rule into each tool's instructions so it holds even when you're not watching.
Where this leaves you
You don't need to learn a new discipline to get AI tools working together. You need the one you already use with people, applied a bit more strictly. Write the one line per tool today and see what changes.
