auto-loop: TDD Autonomy with Checkpoints for AI Agents

Where this fits

You're mid-task on a new feature—say, user login—and your agent starts writing code, then tests fail, then it retries the same approach or loses context when you switch branches. auto-loop turns this into a TDD loop: each acceptance criterion gets a red/green iteration, and every step is checkpointed so you can resume exactly where things stopped.

Why agents benefit

• Agent follows RED/GREEN phases, reducing random flailing by tying each code change to a failing test.
• Checkpoint recovery lets agent resume interrupted work after timeouts, rate limits, or context-length issues without starting over.
• Observability changelog records events like file_write, test_fail, and test_pass, giving you an audit trail of what happened and why.
• Acceptance criteria are parsed into discrete ACs, so the agent can track progress item by item.
• --status, --resume, and --max-iterations flags give you control without hand-holding.

Practical scenarios

Implementing authentication from scratch

You invoke /auto-loop "Implement authentication" with acceptance criteria for login form, JWT generation, and error handling. The agent writes a failing test for the first AC, implements minimal code, and only moves to the next AC when green.

Resuming after an interrupted session

A long build hits a rate limit at iteration 12. Instead of losing the thread, you run /auto-loop --resume and the agent reloads its last checkpoint, identifies the failing test, and continues.

Refactoring with a safety net

You give the agent a legacy module and an AC that existing tests still pass. auto-loop adds one behavior contract at a time, creating a TDD safety net before each refactor step.

Add it to your agent workflow

Install the skill with:

npx skills add claude-world/director-mode-lite --skill auto-loop

Then start a new loop:

/auto-loop "Implement user login

Acceptance Criteria:
- [ ] Login form (email + password)
- [ ] JWT token generation
- [ ] Error handling"

Use --status to inspect progress or --force to clear old state before starting a new task.

Compare before adopting

Look beyond stars. Check whether the agent actually writes meaningful tests first or just patches failing tests. Test checkpoint behavior and changelog completeness across sessions. Compare with your current agent workflow or simple shell scripts, and decide if the TDD contract fits how you review changes.

Why it is worth tracking

With 81 stars and a focused README, auto-loop is young but promising for developers who want more autonomous, observed coding loops. The skill's listing at https://www.openagentskill.com/skills/claude-world-auto-loop makes it easy to watch updates and community feedback. Re-evaluate when your tasks grow longer than one context window or you need guardrails around multi-step agent execution.