Software development
An agent with a working copy of the source of its own: it takes on an issue, submits its change for review, and reads what the automated tests say. A person does the merging.
The situation
Every repository holds work nobody enjoys and that still has to be done: raising a library to a new version, clearing a style warning, reproducing a bug a ticket describes. It is too small to plan for and too large to do in passing. An agent that can check out, build and submit a change takes it on — provided its access ends where the assignment ends.
How a change comes about
From the assigned issue to the failed test run that wakes it once more.
The issue reaches it
A label, a notification from the repository, or a regular look at the list. It takes what is assigned to it — not what it finds interesting.
It fetches the source
The access for that is established only while it works. The working copy lands in the agent's directory, where it is ready again next time. Older copies are removed automatically so the directory does not quietly fill up.
It works, and puts it up
Change, commit, proposed change — with a description of what it did and what it did not. If it could not finish something, it says so explicitly instead of glossing over it.
The tests answer it
If they fail, that wakes it again: it reads the log and makes corrections. The merging stays with the person — the platform has no automation for that and is not meant to.
What it works with
The systems that ship for this area. If one is missing, an MCP server is the quickest route; anyone needing their own wake-up events and finer permissions writes a plugin.
- GitLab
- GitHub
- Jira
- Confluence
- VulnDB
What the platform enforces
What a development agent does not decide on its own.
The merge
It submits, it does not merge. That is not a technical limit but the point at which a person reads the result.
Where it may go
Network access is fixed: the agent reaches the addresses released for it and no others. Pulling a package from just anywhere is therefore not possible in the first place.
What it is missing
If it needs a tool its workplace does not contain, it files a request instead of building a workaround. A person decides on it.
Try it yourself
covey is run by you: one program, a Postgres database, Docker for the workplaces. The instructions are in the documentation.