How Much Should an AI Coding Agent Do on Its Own?

Most AI coding tools give you two choices. Either you click “approve” on every little thing, or you flip a switch that lets the AI do whatever it wants.

The first wears you out. The second makes you nervous. Most people want something in between.

This post breaks coding work into steps, so you can choose which ones the AI does alone.

The steps of a coding job

Almost every coding job goes through the same steps. Here are six that matter for this question:

1. Start. Begin the next stage of the work.
2. Choose an agent. Decide which AI does this task.
3. Retry. If something fails, such as a test, try to fix it.
4. Merge. Bring the finished change into the main version of the project.
5. Push. Send it to the shared copy online so other people and systems can see it.
6. Deploy. Put it live, where real users will use it.

Each costs a different amount if it goes wrong.

Easy to undo, easy to allow

You let a passenger pick the music. You don’t let them grab the wheel on the motorway.

Steps one to three are like picking the music. If an AI starts a stage you didn’t need, you lose a few minutes. If it picks a weaker agent, you can switch. If a retry goes wrong, you throw it away.

Steps four to six are closer to the wheel. A merge changes the main project. A push shares it. A deploy changes what real users see. These are harder to take back.

So a sensible person might say: “Let the AI start, choose and retry. I’ll look at merges myself.” Another person, working on a hobby project with no users, might say: “Let it do everything, I’ll check in the morning.”

Both are reasonable. Their risks differ.

Why one global setting is too blunt

Say a tool has one setting: “autonomous: on or off.”

• Turn it off, and you approve “start the next stage” a hundred times a day. You stop reading them and click yes on everything.
• Turn it on, and the AI may also push and deploy, when you only wanted it to retry a failing test while you sleep.

You need to decide step by step, because the steps carry very different risk. A single switch can’t say “retry freely, but ask before you merge.”

A dial from guided to full auto

A better picture is a dial with a few marked positions. Turn it up and the AI does more by itself. Turn it down and you decide more. Ayut’s home page shows one with five positions:

• Guided: you decide every step.
• Rule-based: the AI starts stages and picks agents by rules you set. You decide the rest, including retries.
• Smart: the AI can also choose agents and retry a failed step, within limits. You still decide on merge, push and deploy.
• Overnight: made for work you leave running while you’re away. Routine steps move forward on their own, a failed step gets a limited number of retries, and merge, push and deploy wait for you.
• Full auto: the AI does the most on its own. You choose which steps it may handle and what the limits are.

Each position is a written list of who decides what. You can read it before you turn the dial.

Full auto is your call

Some people think an AI should never merge, push or deploy without a human. Others think that’s their own decision to make. Ayut’s owner takes the second view. Adults should decide for themselves what Full auto may do. If you want it to merge and push after the checks pass, and you’re happy to accept the risk, that is your project and your decision. If you want it to stop before deploy, you can say so. The tool should make the choice clear and do exactly what you chose.

Full auto still has edges, and you draw them: which projects, which steps, how many retries, how much it may spend.

Every decision should leave a record

There is one thing that should stay the same on every position of the dial: a record.

Whenever a step happens, whether you approved it or the AI did it alone, the tool should write down what happened, when, who or what allowed it, and the result. Like a receipt from a shop.

A record matters for three reasons:

1. You can check the work in the morning without guessing what happened overnight.
2. You can find where it went wrong when something breaks.
3. You can adjust the dial. If the AI handled its retries well, trust it with more. If a merge went badly, turn it down.

Let what you’ve seen in the record decide how much to trust it.

How to choose your own setting

A simple way to start:

1. Mark each of the six steps: “alone”, “alone within limits” or “I approve”.
2. Start with fewer on “alone” than you think you need.
3. Read the record for a week.
4. Move one step at a time.

With real users, you may keep deploy on “I approve” for a long time. For an experiment, you may let it all go. Either is fine.

How Ayut approaches this

Ayut lets you run many AI coding agents on one project, with any AI, from any device, while you stay in charge. The autonomy dial is how you stay in charge: you choose, per step, what runs without you.

To be honest about where things stand: Overnight is the profile that has been proven so far. Guided, Rule-based, Smart and Full auto are designed, and Full auto is not built yet. Watching and approving from a browser works today; running on a desktop host or on your own server is a preview.

If you want to see the dial and the other features, you can Open Ayut.