How to stop babysitting your AI coding agents

If you sit and watch your coding agents work, you are paying twice: once for the agent and once with your own hours. You could be working on the business, on the product or on yourself. Babysitting feels responsible. Mostly it is a sign that no process is catching mistakes for you.

Why people babysit

Two reasons come up again and again. They do not trust the output yet, which is fair. Or they have no checks, so their own eyes are the only check. In both cases, the answer is the same: move the checking into a process that runs without you.

The loop that replaces watching

Tomi's rule is that your job is to define the input and review the output. Both sides of that loop need to be fast and small. If the result is not what you expected, improve the input instead of hovering over the run.

Here is that loop in five steps:

1. Write the task with a finish line. What should work, how to test it, what to follow.
2. Let the agent build in its own space. It opens a pull request when done.
3. Run an independent review. A different agent reads the pull request and flags bugs.
4. Release to staging. Nothing goes to users yet.
5. You check staging and verify QA. Then you roll it out.

You are present at step one and steps four to five. Not in the middle.

How we do it at SupaKid Studios

Tomi is a product manager who also writes software, and he contributes to many projects. He says he merges hundreds of pull requests a week, well over a thousand a month. This is the flow he describes in his video.

It starts with engineering-manager agents. They work like engineering managers: long-lived agents that run other threads for you. Each sub agent has its own cloud computer.

A dev review bot then reviews the code that merged. Boxes on Boat, which are cloud VMs that run CI/CD, handle the app release. It goes to staging. Tomi checks staging, verifies QA, and then it rolls out.

He is also clear about what he still does: he understands the product. When he talks to a client or to his boss, he can explain what is happening in the product and how the code works. He may not know every line, but he can explain how it behaves. That is the job that is left once coding cost drops to near zero.

His point to engineers is that many do not trust their agents enough, or they sit on top of them too much, and that makes them work more than they need to. He also notes the tools are replaceable: Devin, Factory or Claude Code's cloud agents can all fill these roles. The workflow is what matters.

How to build trust without watching

• Make the agent prove it. Devin's docs describe it testing the app end to end after creating a pull request, recording a video of the test and sending it to you. Watching a short recording of the proof beats watching a long run.
• Give tasks checkable outcomes. Devin's own guidance says it works best when it can verify its work, such as by passing CI, unit tests or browser flows.
• Use a manager on a timer. A PM or engineering manager that checks the board every 30 minutes catches stuck work so you do not have to.
• Keep release in steps. Review, merge and rollout are separate. A mistake at one step is caught at the next.

More on the review layer in AI code review and testing agents.

What goes wrong

• Tasks that cannot be checked. "Improve the dashboard" gives the agent nothing to prove. Use "the export button downloads a CSV with these columns".
• Reviewer and writer are the same agent. It will agree with itself.
• No staging. You find out in production.
• Letting go of the product. Hands-off is not mind-off. Review the actual result, especially design and creative work. Nothing here makes a team automatically hands-off.

Zoom out

Once the pipeline is set up, it works beyond code. Tomi points to ad automation and game development as places where an engineer with this workflow can move faster than most. The structure for that is covered in how to structure an AI engineering team.

Next step

If you would rather have the pipeline built and run for you, that is what a managed AI agent team at SupaKid Studios does. The studio team sets up and runs the agents, and you approve what goes out. Book a 30-minute call and we will look at where you are watching today.