What maintenance is
Software maintenance is everything that happens to code after it first works: fixing what breaks, keeping dependencies current, closing security holes, keeping builds and deploys healthy. It is a large share of what running software costs, and it is the work a team is always behind on.
Most of it is not creative work. It is noticing, diagnosing, making a small change and checking it. That is the kind of work an AI system can do, on one condition.
The condition: a signal and a check
An AI system can take on a maintenance job when two things exist:
- A signal that says something needs doing. An error, a red build, a failed deploy, a new entry in a vulnerability list.
- A check that says whether the job is done. A test that now passes, a build that goes green, a scan that comes back clean.
With both, the system knows when to start and when it has finished. Without a check, it is guessing.
What AI can maintain today
These are the jobs Bugstack does. Bugstack does autonomous software maintenance: it finds the work itself and hands back pull requests.
| Job | Signal | Check |
|---|---|---|
| Fix a crash | An uncaught error from your app | A test that reproduced the crash now passes |
| Fix a failing build | A failed run on your main branch | The run goes green |
| Fix a failed deploy | A failed deploy on Render | The next deploy succeeds |
| Upgrade packages | A newer version exists | Your tests pass on the upgrade |
| Patch a vulnerability | A package appears in a vulnerability list | The scan is clean and tests pass |
| Find committed secrets | A secret is in the repository | A person rotates it; Bugstack never does |
| Audit for security problems | You ask for an audit | Each problem is proven before it is reported |
What AI cannot maintain yet
- Wrong answers that look like success. If a page shows the wrong total and returns normally, there is no signal.
- What the product should do. An agent can read what the code does. It cannot know what you meant.
- Judgment about trade-offs. Whether to take a breaking upgrade this week is a decision about your team’s time.
- Problems outside the code. A database that is down or a missing setting on your host can be reported, and a code change will not fix it.
How well it works
We measured it two ways. On SWT-Bench Verified, a public benchmark of 433 real bugs from open-source projects, Bugstack’s tester wrote a test that reproduced 91.5% of them, the top score on the leaderboard. On our own sealed exam of planted bugs, 43 of the 44 fixes Bugstack proposed held.
It is not always right. The answer to that is control: every change is a pull request, and a new repository starts with a person merging each one.
What a person still does
- Decides what the software should do.
- Reviews and merges changes, until they choose to let proven ones merge on their own.
- Handles what has no signal.
- Rotates secrets and fixes problems outside the code.
The work that moves to the machine is the noticing, the diagnosing and the routine fixing. That is a large share of maintenance, and it is the share nobody enjoys.
Questions
Can AI maintain software without any human?
For the jobs above, it can do the work without being asked. Whether its changes merge without a person is your choice, per repository. On Bugstack, new repositories start on Review, where you merge everything.
Will AI replace the engineers who maintain software?
It takes over routine upkeep: the crash nobody has picked up, the upgrade nobody has time for. Deciding what to build, and judging the changes, stays with people.
Can AI maintain code that AI wrote?
Yes. Code is code. An app written with an AI assistant breaks and ages like any other, and often has fewer tests, which makes a system that watches it more useful.