How I work with AI agents
Card Loop Pipeline
This is how I iterate with AI agents: each card moves through seven connected stages. The loop alternates between autonomous phases and human decision points.
- 01. Triage — Pick up + classify
A card is picked up from the backlog in To Do. The loop triggers automatically when a card with the agent:loop label enters the right state. Classifies task type and assigns priority.
- 02. card-start — Worktree + branch
Creates an isolated worktree with its own branch from the Jira card. Resolves the repo and base branch from the card's labels. Ready to work without touching the main checkout.
- 03. SDD auto — Explore → Spec → Design → Tasks
Full 5-phase pipeline: Explore (context), Propose (intent), Spec (requirements + scenarios), Design (technical architecture), Tasks (breakdown). All generated by agents before touching code.
Human-in-the-loop - 04. card-review — Pre-PR 4R review
Pre-PR review using the four 4R lenses (Requirements, Reliability, Readability, Risk). Report-only: catches issues before opening the PR. Optionally simplifies code if requested.
- 05. card-pr — Open PR
Pushes the branch and opens the GitHub PR. Does not merge — leaves the diff ready for the adversarial review in the next step.
- 06. Judgment Day — Dual blind review
Blind adversarial review: two agents review the diff independently, compare findings, and deliver a verdict. If issues are found, it loops back to SDD for fixes + re-judgment (max 2 rounds).
Human-in-the-loop - 07. Partners / Slack — Code Review + notify
The card moves to Partners/Code Review in Jira and a single Slack notice goes to the backend or frontend channel. Merge is human — the loop never auto-merges.
Legend
How it terminates
The pipeline ends when card-pr publishes the PR, Judgment Day returns a clean verdict, and the Partners channel receives the notification. No auto-merge: the final decision is always human.
Want to see it in action on a real project?
Browse projects