Agent Architecture and Orchestration
Loop Engineering — Stack
Loop engineering › Stack
Legacy handbook baseline: May 2026. Source-verified additions reviewed 2026-10-13. The loop terminology below includes practitioner material published after the baseline.
Choosing the outer controller
| Mechanism | Appropriate work | What it does not decide for you |
|---|---|---|
| Temporal Schedules and workflow executions | Recurring work with recorded execution and configurable overlap/catchup behavior | Work eligibility, acceptance evidence, and item-level attempt limits |
| GitHub Actions triggers | Repository events and periodic code-related tasks | Whether a scheduled dispatch is timely enough and which business item is already claimed |
| A queue consumer with a durable work table | Event-driven backlog and explicit ownership | Deduplication, stale-event handling, and uncertainty after worker loss |
| An in-process supervisor | Short local experiments | Persistence and unattended availability after the parent application exits |
Admission before autonomy
Start with one work class, a fixed acceptance checker, and a bounded attempt policy. Verify what happens to an item after a worker is replaced, after a trigger repeats, and after a write succeeds but its response is lost. Capture controller, harness, and checker revisions separately. Automatic dispatch is useful only when the system can distinguish verified work from busy runs and blocked prerequisites.
The Loop Engineering essay motivates the cross-run controller, while Own the Outer Loop discusses the continuing human responsibility. Neither essay establishes universal product requirements or justifies adopting every automation primitive it names.