Class 7 · 7.4

Ralph Loops: autonomous coding on autopilot

Put the agent in a loop so that every time it thinks it's finished, it goes again. It's how an eight-minute task becomes an overnight build — where the idea came from, how to run it, and when it's the wrong tool.

🎯 Goal: understand autonomous looping⏱ Read: 10 min🧭 Type: hands-on

The finale for the day is genuinely a little crazy: put the agent in a loop so that whenever it thinks it's finished, it's nudged to start again and do more. This is a Ralph loop — and it turns an eight-minute task into an eight-hour one.

01The idea: a loop around the loop

The agent already runs an inner loop — think, act, observe, repeat — until it decides the job is done. A Ralph loop wraps an outer loop around that: on "done," it re-prompts Claude (carefully, to keep the context from ballooning) to repeat the task from wherever it is and do it better, over and over, for a set number of iterations.

Where it came from

The Ralph loop is the brainchild of Australian engineer Geoffrey Huntley, named in homage to Ralph Wiggum from The Simpsons — the endlessly optimistic character who just keeps going. Couple a Ralph loop with YOLO and you can set something running for hours, each pass building on the last.

02Running one — via a plugin

Ralph loops ship as an official Anthropic plugin (we cover plugins properly next class). Install it, then call its command with a prompt and an iteration count.

Claude Code

# install the official plugin
/plugin install ralph-loop@claude-plugins-official

# then kick it off: prompt + how many times to loop
/ralph-loop:ralph-loop "Significantly improve this project. Add users and
  user management, multiple Kanban boards per user, and other
  features to build a comprehensive PM app. Test thoroughly and
  keep strong coverage and integration tests." --max-iterations 10

It writes a ralph-loop-local.md file inside .claude and sets off. Because the demo ran it without full YOLO, it paused for approvals — always press 2 for things you're happy with, so you keep a hand on the wheel and avoid anything destructive.

Ralph + YOLO belongs in a sandbox

Running a Ralph loop on plain YOLO with no sandbox is not a best practice — it's a lot of unattended, unapproved action. Do it only inside a sandbox (covered next week), or stay present and approve as it goes.

03What an hour of looping produced

The demo ran for just under an hour and changed ~1,300 lines of code — compacting itself along the way. The result was a real step-change: proper authentication with account creation, per-user multiple Kanban boards you can flip between, persistence across logout/login. It had crossed from "toy" toward "platform you could actually ship."

Before

Single board

One hard-coded board, a demo login, no real persistence.

≈1 hour later

Multi-board app

Sign-up, auth, per-user boards, switch between them, data that survives logout.

The point

Autonomy scales

Bigger iteration counts = overnight builds. Tens of thousands of lines, unattended.

04When to reach for Ralph loops — and when not to

"Why not always use them?" Because you trade control for reach. The answer is the same one that runs through this whole course: match the tool to the job.

Reach for a Ralph loop	Stick with a regimented process
Prototypes, MVPs, "out-there" experiments	Work needing predictable, reliable outcomes
You want maximum autonomy and reach	You want to play a primary, hands-on role
Overnight "just keep building" sessions (sandboxed)	Production code you'll own line by line

A teaser for what's coming

Ralph loops are a first taste of multi-agent, autonomous coding. Over the next week and a half we'll make this robust — with sub-agents in defined roles, testing baked into the process, and sandboxes for safety. This was the fun, slightly reckless preview.

✓ Key takeaways

A Ralph loop wraps an outer "not done — go again" loop around the agent's normal loop, for N iterations.
It's an official plugin: /plugin install ralph-loop@claude-plugins-official, then /ralph-loop:ralph-loop "<prompt>" --max-iterations N.
Without YOLO you approve as it goes (press 2); with YOLO, only inside a sandbox.
An hour turned a single-board toy into a multi-user, multi-board app — scale the iterations for overnight builds.
Great for prototypes and MVPs; use a regimented process when you need predictable, ownable results.

That's a wrap on Class 7 — 47% of the way through

You drove the commands, learned the shortcut keys, mastered sessions / checkpoints / Git, ran real YOLO, and let a Ralph loop build for you. Next is an inflection point: MCP, Skills & Plugins — the three building blocks that take your Claude Code expertise to a new level.