Autonomous Improvement Loop

A recurring scheduled pattern where an agent improves code targets without per-cycle human direction. The core loop is Observe → Orient → Decide → Act → Repeat, plus one discipline: finish in-flight work before starting new work.

This is distinct from one-shot task execution — it is a persistent process that compounds small improvements over many cycles.

Setup (once, then reuse)

Fix three decisions before the first cycle; record them where the schedule can see them:

1. Target scope — a single repo, a round-robin portfolio list, or a priority queue.
2. Delivery vehicle — draft PR, issue, direct commit (only with explicit standing approval), or report. Draft PRs are the default: reviewable, gated, reversible.
3. Identity convention — how loop output is recognized later (branch prefix, label, title tag). Choose once; the references below illustrate with branch autopr/improve/ + label autopr.

Done when: scope, vehicle, and convention are written down and visible to every scheduled run.

When to Use