Atlas Opsbook
A coding agent can write its own operations manual. How does it decide which lessons are worth keeping?
When shipping a package, install the built artefact in a clean environment and import it.
Scope: package releases
Where should the rule apply?
A package passed its source tests, but the delivered archive could not be imported. The agent proposed checking every release. Compare that draft with the narrower rule it learned from four practice cases.
| Task | Manual asks for | Outcome |
|---|---|---|
| Ship a macOS package | No check | Missed check |
| Ship a FreeBSD package | No check | Missed check |
| Publish release notes | No check | Correct |
| Inspect package metadata | No check | Correct |
The empty manual misses both package checks. These are recorded synthetic cases. An independent live-agent comparison is still open.
Give your agent a manual.
Run the offline exercise with Git and Python 3.9 or later. The archive trial builds a broken package, catches the import failure, then tests the repair.
Built by Albert Tan Sin Hon. Atlas Opsbook works alongside Lauren Tan’s pstack. The original engineering workflows remain available through a pinned upstream distribution.
git clone https://github.com/Albertito-AI/atlas-opsbook.git
cd atlas-opsbook
python3 opsbook.py demo
python3 examples/package_trial.py