An agent loop with write access. Almost everything waits for my review.

An agent that can open a repository, propose a fix and merge it is easy to build. The hard part is letting it run: an agent that gets it wrong while holding write access damages work, and nobody notices until the following week.

The Command Center runs locally, with no server and no public URL. It scans the monitored repositories and the Obsidian vault, ranks them by an attention score, and routes every agent-proposed action into a review queue with gates, an audit log and a kill switch. Nearly all of them stop at human review, and everything lands in the log.

Autonomy is scoped in policy to the Command Center's own repository. Every other monitored repository is proposal-only, so a bad agent decision can only damage my own tooling. The degraded path is stated outright: every failure falls back to review, never to merge.

What is allowed through without me is a short, closed list: documentation drift, lint errors (three at most), patch or minor dependency updates (two at most). Everything else waits.

Seven weeks of logs: 59 autopilot runs, one of which failed. The rest of the queue I handled by hand, one proposal at a time.

The logs keep every decision, line by line, which makes that restraint checkable without taking my word for it.