What an AI coding agent does on its own, and what it still needs you for

“AI coding agent” and “autonomous AI agent” get used as if the model were the whole thing. It is not. The model writes. The agent is the loop around the model, and the loop is what you are actually paying for.

The loop, in one paragraph

You describe a job. The agent writes code, runs it, opens the result in a real browser, and runs a set of acceptance checks: does it fit a phone screen, are there script errors, is the arithmetic right, is the technical SEO complete. If something fails, the failure is fed back as text and the model tries again. When the checks pass, the job is done. Every one of those tries is a turn, and every turn is a charge.

That is why the cheapest model won most of our benchmark: it finished page builds in one turn. A model that costs a third as much per turn but needs two turns loses. Turns to done is the number.

What it genuinely does without you

• Reads its own errors. A type error, a missing import, a failing test: the agent sees the message and fixes the cause, usually on the next turn.
• Checks the result like a person would. Not “the build succeeded” but “the page renders, the button works, the price adds up.”
• Escalates when it is out of its depth. On Balanced, a failed job asks the next model up the ladder, one rung at a time, rather than jumping to the most expensive one.
• Ships. The code lands in a git repository you own, and the app runs at a real address with hosting included.

The failure that cost a customer twice

Here is the case that changed how we build this. A customer’s app went down. The platform’s deploy step hit a Docker mount error, a bug on our side that no code change could fix. The agent was told only “the build failed.” It read that as its own work being wrong and started fixing correct TypeScript. When that did not help, it escalated: five models, one after another, rewriting code that was never broken, spending 26 cents of a 70-cent balance. The app stayed down the whole time.

Nothing in that loop was stupid. Every step was the right response to the signal it was given. The signal was wrong. So the fix was not a smarter model; it was teaching the platform to recognise the handful of errors that cannot be caused by code and to stop the loop instead of feeding it. An autonomous agent is exactly as good as the truth of what it is told.

The three things it still needs from you

1. A budget it cannot cross. Not a suggestion. A number, enforced before the next step, not after the bill. Every job here shows its running cost and stops at the cap.
2. A yes before anything irreversible. Sending an email, charging a card, deleting data, publishing to a domain. The agent proposes; a person approves. We put that approval on a home-screen widget because the moment matters more than the screen you happen to be on.
3. The job, stated as an outcome. “Bookings page that stores bookings” is checkable. “Make it nice” is not. The acceptance checks are only as good as the sentence they came from.

Where to start

The cheapest first job is one where the agent changes something rather than inventing it. Pick one of the 53 templates, tell the agent whose business it now is, and watch the first turn cost cents. From there, every job you give it is a sentence, a price, and a result you can open.

Which model for which job What it costs