Learn › After you ship

Who maintains code that AI wrote?

You do. When an AI assistant writes code, the responsibility for it stays with whoever ships it. The assistant does not watch your live app, does not learn that a package became vulnerable last week, and is not there at night when a deploy fails. Somebody has to be, and there are three ways to arrange that.

Why the question is new

Until recently, the person who wrote code could also maintain it. They understood it, so when it broke they knew where to look.

AI changed how fast code gets written without changing how it ages. A team now ships far more code than it could write by hand, and one person can ship an app they could not have written at all. The amount of code needing upkeep went up. The number of people who understand it did not.

What maintaining it involves

JobHow often it comes up
Fixing crashes real users hitRegularly, for an app in use
Fixing a failed build or deployWhenever something upstream changes
Upgrading packagesContinuously; new versions come out every day
Patching security flawsWhenever one is published in something you use
Noticing any of the aboveAlways

None of this is building. It is the work that keeps what you built working.

Three ways to get it done

You, with your assistant. When something breaks, you paste the error in and ask for a fix. This works for what you notice. It depends on you noticing, and on being available.

A person you hire. A developer or an agency takes over upkeep. You get judgment, and you pay for the time it takes someone to learn code no person wrote.

A system built for it. Autonomous software maintenance is upkeep done by AI agents that find the work themselves. This is what Bugstack does: it watches your app and your repository, and when something breaks or goes out of date, it opens a pull request with the fix.

These combine. Many teams keep a person for decisions and let a system handle routine upkeep.

Why AI maintaining AI-written code is not circular

It sounds odd to have AI look after code AI wrote. It works because maintenance has something writing does not: a check.

When you ask an assistant to build a feature, nothing can confirm it built the right thing. When Bugstack fixes a crash, a test fails before the fix and passes after it. The result can be verified whoever wrote the original code.

On a public benchmark of 433 real bugs, Bugstack’s tester reproduced 91.5% with a failing test, the top score on the SWT-Bench Verified leaderboard.

What stays yours

  • Deciding what the app should do.
  • Accepting changes. Bugstack opens pull requests, and on a new repository nothing merges until you merge it.
  • Anything with no signal, like a page that shows the wrong answer without failing.
  • Rotating secrets and changing settings on your host.

Questions

Is the AI company responsible if code its model wrote breaks?

The tool that wrote the code does not run your app or watch it. Responsibility for what you ship sits with you, the same as with code from a library or a contractor.

Can the same assistant that wrote my app maintain it?

It can fix what you bring to it. It cannot notice problems on its own, because it is not connected to your running app or your repository between conversations.

Do I need a developer if I use a maintenance tool?

For routine upkeep, the tool does the work and you review it. For decisions about what to build and how, you still want a person’s judgment, yours or someone else’s.

See it work on your own code.

$100 a month

Includes $100 of usage credit every month. Start with a 14-day trial: $50 of credit and a first security audit on us. A card is required.