- A team member processed 108 PRs in eight days using an AI agent that picks tasks, writes code, opens PRs, and handles feedback autonomously, a significant increase from their previous 5-10 PRs per week.
- Key enablers include a small, high-trust team without mandatory human PR reviews; alignment happens through daily conversations, not review tabs, and human review shifts to testing outcomes instead of line-by-line code checks.
- Instead of continuous deployment, the author uses CI plus preview/stage environments, batching work for release to limit potential issues and reduce regression testing to once or twice daily.
- The process is called 'loop engineering,' with three layers: a protocol (rules in a markdown file), a loop (reads board, runs commands, dispatches work), and a worker (writes code in isolated worktrees and reports back).
- The task board uses individual markdown files per task to avoid collisions, with Obsidian's Bases for structure, and the loop is the only writer to prevent file corruption.
- The loop operates self-paced without an interval, stops automatically when no tasks can proceed without human input, and includes a memory file with hard cap for tracking codebase facts and preferences.
- Human effort has shifted to spec work—defining what goes in the queue and verifying outcomes—making the role feel like a mix of product manager and QA.
- Future improvements include automating task creation from issue exploration and automating QA stages like accessibility and regression checks.
- To get started, the author recommends defining 'done' (e.g., task completed, green CI, approved AI review) and then working outward from there.