AbjakGovernance for AI-assisted software work
The agents write. You approve.
Manifest assigns the work, keeps it off the live branch, watches the checks, and explains the change in plain English. A person decides what ships.
01 / The problem
The problem
- 01A chat can say the work is done when nothing was pushed.
- 02The change can land on the live branch.
- 03A green check can be an old failure that was already there.
- 04A cost figure can look like a bill when it is only an estimate.
02 / What happens, in order
What happens, in order
- 01
Suggest.
Manifest proposes a task. Approve creates it. Skip throws it away.
- 02
Assign.
A person sends it. Manifest opens a GitHub issue with the acceptance criteria.
- 03
Isolate.
The agent works on a working branch. It does not write the live branch.
- 04
Prove.
GitHub checks come back onto the task. A model's own summary cannot mark the task done.
- 05
Explain.
The change is described in plain English, labelled as advice when a model wrote it.
- 06
Approve.
A person opens the pull request and a person merges it.
03 / What stays with a person
What stays with a person
- Approve or skip
- Click Create PR
- Merge
- The keys, stored in Integrations, not in the page
04 / One real example
One real example
In progress
- Client work
- CSJ OpsHub, one repo, working branch
manifest/csj-v1, live branchmain. - Task
- fix three helper scripts.
- What happened
- the agent pushed a branch, Rudy opened the pull request, and the checks came back onto the task.
05 / Who it's for
Who it's for
- 01an owner who runs several systems and doesn't read code
- 02a project tied to one GitHub repo
- 03a team that already uses an agent and wants a record of what was assigned, checked and approved