Errors fixed by agents, checked by people
Every error our platform hits in production becomes a task on the right project. A coding agent works out the cause, the size and the severity, fixes it on a separate copy of the code and proposes the change. Small, safe fixes go in on their own once every check passes; anything sensitive or large waits for a named reviewer.
On 12 October 2026, each from an error seen in production.
Of code, in at most 3 files, with a test and every check passed.
New errors in that window pause automatic merging on the project.
Since 12 October 2026. Before, every 5 minutes.
Most errors never take a site down. They come and go, and they still hurt the people using the software. Each one has to be noticed, traced to its cause, fixed and checked, and that queue grows faster than a small team clears it.
Our monitoring already caught them. The slow part was everything after: reading the error, finding the code, writing the fix and testing it. We wanted agents to do that part, and people to keep the decisions that carry risk.
We built it into Desk, the platform we run SEIRO on, and pointed it at Desk first.
Step by step.
- 01
An error in production becomes a task on the project it belongs to. When the fix is in, the task closes and the error is marked as resolved.
- 02
A cloud coding agent triages it first, without changing any code: the probable cause, the files involved, the size, the severity and how confident it is. The result goes on the task.
- 03
If the triage allows it, the agent builds the fix on its own branch, a separate copy of the code, adds a test for the bug and proposes the change for review.
- 04
A fix goes in on its own only when it is small (at most 30 lines of code in 3 files), touches no sensitive area, includes a test and passes every required check. Changes to money, sign-in, permissions, data or deployment always go to a person.
- 05
Large changes, low confidence and failed attempts go to the project’s named reviewer, with a live preview of the fix where the project has one. A failed fix gets one retry first.
- 06
If new errors appear within 6 hours of an agent’s change going in, automatic merging pauses on that project until a person resumes it.
- 07
Alongside it, uptime checks run every 60 seconds on the sites we look after: whether each page answers, its security certificate, its domain expiry and its speed. When a site goes down, the alert email carries an AI diagnosis of the likely cause.
Built with Error monitoring · Task board in Desk · Cursor cloud agents · Code review and automatic checks · Uptime checks every 60 seconds · AI diagnosis of outages
The first fixes were small and real: pages and background jobs that failed in situations nobody had tested. Each one came from an error our platform had hit in production.
Triage before code pays off. On one of the first errors, triage took about 3 minutes and rated the fix large, with low confidence, so it waited for a person instead of guessing.
What counts as sensitive depends on the project. One fixed list was too blunt, so each project keeps its own list, on top of a set of areas that are always reviewed by a person.
What we decided.
Adopted on our own platform since 12 October 2026. Client projects are next, under the same rules: fixes only during development, warranty or a maintenance plan, each project with its own reviewer.
Agents that triage and fix errors fit any software that runs in production and has tests. For your systems, that can mean:
- Errors in your app turned into tasks with a probable cause and a size before anyone opens the code.
- Small fixes proposed with a test that proves the bug is gone.
- A named person approves anything that touches money, data or sign-in.
- Your site checked every minute, with an AI diagnosis in each outage alert.
- Changes to sensitive areas never go in on their own. A person always reviews them.
- Every run has a time and cost limit. A run that reaches it stops and goes to a person.
- Only errors that reach monitoring become tasks. Bugs nobody has triggered still need tests and people.
- So far the fixes run on our own platform. Client projects are next.
Bring it to client projects, and measure the accuracy and cost of triage over the first month to tune what may go in on its own.
Can an AI agent fix production bugs on its own?
Small ones, within limits. A fix goes in without a person only when it is small, has a test, passes every check and touches nothing sensitive. Everything else is proposed for a named reviewer to approve.
What if an agent’s fix makes things worse?
If new errors appear within 6 hours of an agent’s change going in, automatic merging stops on that project until a person has looked and resumes it.
How is this different from the site that came back in 9 minutes?
That agent works on servers: outages, updates and settings. These agents work on the code of an app, fix the errors behind it and propose the change for review.