AI Giants Recap: What Everyone Gets Wrong About The Ralph Loop

“Any company out there that’s raised a large amount of venture capital who has adopted a BSL license or some bespoke license, none of that matters anymore. That’s not a safety net. It’s gone.”

That line from Geoffrey Huntley, goat farmer, long time open source contributor, and creator of the Ralph Loop, hangs over this AI Giants episode.

Host and Codacy CEO Jaime Jorge spends the hour pulling on that thread. What happens when a few people with a loop can clone the behavior of your product, and why that should keep founders and engineering leaders up at night.

P.S. Catch Jaime’s recap of the episode on X here

What is the Ralph Loop?

The Ralph Loop is not a SaaS product or a classic “AI agent”. It is a bash loop that spins up a fresh context window on every iteration and lets the model do exactly one task each time.

The loop feeds in specifications that act as a lookup table and framing device for the work to be done. That way, the model never relies on a big, stale context where compaction can quietly erase the important parts and invent new ones.

TLDR For Engineering Leaders

• Software development is over. Software engineering is not. Geoff puts the cost of development work at 10.42 dollars an hour, lower than minimum wage, and says anyone can do it while they sleep. The value shifts to people who know how to design systems, reason about failure, and think about security.
• Licenses are not a moat anymore. By running Ralph in reverse over source code, product docs, and marketing material, Geoff was able to recreate HashiCorp Nomad’s feature set as a clean room specification without copying a line of code.
• Specs are the real asset. Institutional knowledge lives in specification files that every loop reads. Code becomes something you throw back on the “pottery wheel” whenever you need a refactor, a migration, or a new feature.
• Most “agent” tools are just loops with branding. Cursor and Windsurf are, in Geoff’s words, loops that automate the copy paste dance between IDE and model. The underlying pattern looks the same.
• AI magnifies whatever skill you already have. Someone with a deep security mindset plus loops becomes extremely effective. A junior without guidance becomes a very fast way to embed bugs and vulnerabilities.
• The bottom rung of the ladder is disappearing. Geoff describes junior doors closing, reopening only after teams rethought apprenticeship for an AI native world. In the meantime, a cohort of juniors has been practicing loops for months and will happily work at a fraction of senior rates.
• Languages are being re-ranked by how well agents can work with them. Rust, Python with Pydantic, and TypeScript with Effect.js sit at the top of his list. Java and .NET sit at the bottom because DLLs break the simple “ripgrep the codebase” pattern these loops rely on.
• Guardrails on AI-generated code are no longer optional. If loops can touch large parts of your system without a human in the loop, you need automated checks that enforce security and quality standards on every generated line.

How It Started: A Nine-Year-Old and Factorio

The origin story of Ralph is not a lab or a white paper. It is Geoff playing Factorio with his son while using Cursor on the side.

He kept doing the same steps by hand. His son watched, then said, “Dad, you seem to be doing a lot of things by hand, but it is the exact same way every time. Why don’t you just put it in a loop?”

Geoff tried it. “Holy crap.”

The models were still rough. He describes an earlier Claude Sonnet model as “a hyperactive squirrel armed with a kitchen knife” that will stab you if you look away. You can get work out of it, but only if you know how to tame it.

Development vs. Engineering

Geoff splits the profession in two: development, which he compares to “JIRA ticket monkey” work, and engineering, which is about shaping systems and understanding how they can fail.

With Ralph style loops, the development side is cheap and abundant. You can get “weeks worth of work of potentially multiple co workers in a couple hours” of looping. The part that remains scarce is engineering judgment about requirements, failure modes, and security.

"You could probably say that software development is over now because anyone can be a software developer."

Ralph In Reverse: The Nomad Experiment

To test how far this could go, Geoff did not start with a toy app. He started with HashiCorp Nomad.

He used Ralph to turn the Nomad codebase into specifications, then realized he was missing about twenty percent of the behavior because the enterprise features were not in the public source. To fill that gap, he ran Ralph over Nomad’s product guides and marketing material. That was enough to plug the missing pieces.

That experiment convinced him that for many funded software companies, the perceived moat from licenses and hidden enterprise code is already gone.

The Context Window is Clay on a Pottery Wheel

Earlier in his career, Geoff thought in Lego metaphors: build libraries as solid bricks and assemble systems from known good parts. If something fails, fix the brick and keep the tower standing.

Now he thinks in clay. The context window is a lump of clay on a pottery wheel. He sits down with the model, feeds in the specs, and shapes the clay through a back and forth conversation about design choices like ABAC versus RBAC. At the end of about thirty minutes, he has a well-formed specification he can hand to Ralph.

If something is off, he throws the code back on the wheel. One loop to clean up duplicated SQL. Another to add internationalization. Another to apply a repository pattern. Over time, he imagines fleets of specialized “Roombas” that keep codebases tidy and secure, while other loops focus on shipping features.

Specs As The Place Where Knowledge Lives

Because every loop starts from the specs, Geoff is comfortable with each iteration “forgetting” the rest of the conversation. That is not a bug in his view. It is how you avoid compaction silently dropping the important parts of the context.

“My institutional knowledge is in the specifications file”

Each spec links to others. Together they form the reference frame the loop uses to understand the domain and the patterns that should be followed.

This has a direct impact on language and framework choices. Geoff ranks languages by how well agents can search and transform them with tools like ripgrep and how strongly the type system can reject bad generations. That is why Rust, TypeScript with Effect.js, and Python with Pydantic are at the top of his list, and why Java and .NET fall at the bottom.

Engineering Careers And The “Hard Line” Ahead

Geoff talks openly about shutting the door on junior roles for a while at a previous company until they could figure out how to assess AI native engineers. During that time, a cohort of juniors spent six or seven months experimenting with loops, learning tricks like Tmux based workflows and GitHub CLI driven CI fixes.

Those juniors will now work for roughly a quarter of a senior’s salary. Mid career engineers who are only just opening an AI assistant after Claude Opus landed are, in his view, in a much more precarious position than they think.

His advice to the mid career engineer with a mortgage and kids is not to carve out a 60 hour learning block. It is to pick up the “new computer” in thirty minute sessions, keep experimenting, and notice when you are too excited to sleep because of what you can now build. That is the signal you actually get it.

Why AI Feels Like Picking Up A New Computer

Geoff keeps coming back to the feeling that we are all picking up our first computer again.

For decades, operating systems, CLIs, and languages were designed around human operators: type into a TTY, edit environment variables by hand, and manage breaking changes carefully so you do not break people’s mental models. Loops make a different set of tradeoffs possible. You can ship breaking language changes alongside automated migration loops. You can rebuild tools around what robots need rather than what humans type.

That is why, he says, some of the people who shaped computer science are coming back out of retirement. They see a genuinely new machine to tinker with.

When Open Source Stops Making Sense

One of the more uncomfortable parts of the conversation is Geoff’s stance on open source. The same person whose family home was funded in part by open source donations now avoids pulling in open source dependencies where he can.

His reasoning is practical. Open source brings humans back into the loop. If a library has a bug, you file an issue or a pull request and wait. For Ralph style workflows, any need to call a human means you lose the properties you are aiming for. It is often easier to generate the code you need and own the responsibility for it.

There are clear exceptions. You do not generate your own cryptography primitives and trust them blindly. You still need specialists who can vouch for security critical code. But for large parts of the stack, Geoff has decided that generation beats dependency.

Code security and code quality in the Ralph Loop Era

When Geoff describes loops that can reshape entire codebases in hours, or junior engineers working at a quarter of senior salaries while generating weeks of code overnight, you need the guardrails in place before things go sideways.

This is where automated code quality and security checks stop being nice-to-have and become load-bearing infrastructure.

If a loop can touch your authentication layer, rewrite your database access patterns, or refactor your API without a human reviewing every line, you need something watching that code before it ships. You need to know that every generated function follows your security standards, that no one is accidentally embedding SQL injection, and that the "pottery wheel" didn't quietly introduce a vulnerability while reshaping your codebase.

Geoff's right that software development is cheap now. But software engineering (the part that understands failure modes, thinks about security, and enforces standards across systems) just became exponentially more valuable. The companies that survive this shift will be the ones that figured out how to amplify engineering judgment with automation, not replace it.

Codacy sits right at that intersection. Because in a world where code is clay and anyone can spin the wheel, the only moat left is knowing the code is safe before it hardens.

AI Giants is Codacy's podcast series featuring conversations with leaders building the future of AI coding. Watch the full episode with Geoffrey Huntley on YouTube.