Using Backlogs To Queue Tasks For AFK Agents.mp4

Commits:

# see solution — prompt updated to fetch GitHub issues with task selection
pnpm reset 07.06.01
pnpm cherry-pick 07.06.01

Using Backlogs To Queue Tasks For AFK Agents

Your AFK agent is currently stuck in a bind - it needs you to manually point it at a plan and a PRD every time you want it to work. That's not really "autonomous" at all, is it?

What if instead, your agent could pull tasks from a queue and decide which one to work on next? This creates a wonderful feedback loop: your agent writes code, humans review it, new issues get created or reviewed, and your agent picks up the next task automatically.

You can use your existing issue tracker (like GitHub) as the source of truth. Your agent becomes a true autonomous worker.

How It Works

The updated prompt in .sandcastle/prompt.md now fetches open GitHub issues instead of pointing to a static plan and PRD. It runs this command:

gh issue list --state open --json number,title,body,comments

The agent uses a task selection prompt that prioritizes work in this order:

  1. Critical bug fixes - Broken CI or broken app functionality
  2. Development infrastructure - Tests, types, and dev scripts (important precursors to building features)
  3. Tracer bullets - Small end-to-end slices of new features that validate your approach
  4. Polish and quick wins - Improvements and easy wins
  5. Refactors - Code cleanup and restructuring

This prioritization setup works well for most projects up to around 20-30 open tasks. You'll want to adjust it carefully based on what stage your project is at.

Task prioritization prompt showing the five levels of priority

What Makes Tracer Bullets Special

Tracer bullets are small slices of functionality that go through all layers of your system. They let you test and validate your approach early.