Coding agent loops

Coding Agent Loops — Atlas

Run autonomous coding tasks with checkpoints and sprint integration.

What This Is

A structured loop for executing coding tasks from the sprint board end-to-end: read task → plan → implement → verify → mark done → next task.

Atlas tracks progress at each checkpoint and surfaces blockers instead of silently failing.

Starting a Loop

"Run coding loop for [task]" / "Autonomously implement [task]" / "Work through the sprint"

Pre-Loop Checklist

1. Read memory/sprint.md — confirm task exists and is P1/P2
2. Confirm task is not blocked (memory/blocked.md)
3. State the task clearly: "Starting: [task title]. Success = [definition of done]."
4. Ask: "Proceed?" — do not start without confirmation

The Loop

LOOP:
  1. READ   — understand current state of the codebase for this task
  2. PLAN   — state what you'll change in ≤5 bullet points
  3. CODE   — implement one logical unit at a time
  4. VERIFY — run tests / build / lint; check it works
  5. CHECKPOINT — report status; ask to continue or stop
  6. If done: mark task done in sprint.md, move to done.md
  7. If blocked: write to blocked.md, stop and report
  8. If complete: propose next P1 task from sprint

Checkpoint Format

After each logical unit of work:

Checkpoint — [task title]

Done so far:
- [x] Created auth module
- [x] Added JWT validation

Next:
- [ ] Wire up to user routes
- [ ] Add tests

Blockers: none / [describe blocker]

Continue? (yes / stop / change direction)

If user says stop: write current state to sprint.md, save progress note. If user says change direction: re-plan from current state.

Blocker Handling

If the loop hits a blocker:

1. Stop immediately — do not push through uncertainty
2. Write to memory/blocked.md:

   - [~] [P1] Task title | blocker: [specific reason] | date: YYYY-MM-DD
   

3. Report: "Blocked on [task]. Reason: [specific reason]. Options: [A] / [B] / [C]."
4. Wait for user decision before continuing

Loop Rules

1. One task at a time. Never start task N+1 while task N is in progress.
2. Checkpoint every logical unit — not every line, not every hour. After each coherent piece of work.
3. Verify before marking done. Tests pass or build succeeds = done. Nothing else.
4. Surface blockers immediately. No "I'll work around it" — stop and report.
5. State your plan before coding. No silent 200-line PRs. Plan first, code second.

Sprint Integration

On task completion:

# Atlas updates sprint.md automatically:
# - [x] [P1] Task title | completed: YYYY-MM-DD | sprint: N

Then proposes the next P1 from memory/sprint.md.

If sprint is complete (all tasks done): trigger sprint review via sprint-planner skill.