Huntley's take on Ralph loop: why Anthropic and Vercel approaches shouldn't be conflated

The term Ralph loop quickly became an umbrella for very different agentic architectures. Recent analysis shows that under one name today there are mixed at least five patterns — from a simple model restart loop to systems where an agent changes its own instructions and artifacts between iterations.

Why the dispute arose

The author of the article starts with a simple question: what should be considered a true Ralph loop. A quick search through public threads, READMEs and blogs didn't clarify the picture, but only added confusion. Some call Ralph a simple external loop that runs the same prompt again, others call it a scheme with a separate verifier, still others call it an almost self-evolving agent.

As a result, under the same name, people began discussing constructions that by design and risks are close only at a distance. To bring order, the author suggests looking not at the brand, but at architectural characteristics. The key questions here are: where is the verifier located, who acts as the oracle, where do the completion criteria live, and what exactly is carried over between attempts.

A separate line is the right to mutation: can an agent only change the working plan, or is it allowed to rewrite checks, specifications, and even its own system prompt. It is precisely this choice that affects the safety of the entire scheme.

Five versions of Ralph

The article presents a working taxonomy of five patterns that today are most often hidden under the name Ralph. They are indeed similar at the facade level: everywhere there is a loop, an attempt to fight context rot, success criteria, and some kind of verification mechanism. But as soon as you look deeper, it turns out that in some systems the model itself decides when to stop, while in others this right is taken outside, and between iterations what changes is not only the output, but also the working artifacts.

• Same-prompt Ralph in the spirit of Anthropic: the same prompt runs again and again until the model itself decides to say DONE, and the external loop only catches the stop signal.
• External verifier Ralph in the Vercel model: the external verifyCompletion is already separated from the internal tool loop, but the initiative to exit an attempt still remains with the model itself.
• Artifact-evolving Ralph in Geoffrey Huntley's original version: between iterations, not only logs change, but also useful artifacts like a plan, working rules, and accumulated lessons.
• Artifact-evolving Ralph with external verifier: a stricter variant where artifacts evolve, but success criteria are fixed, and an external validator can roll back unauthorized changes.
• Self-evolving agent: already almost a separate class in which multiple agents can analyze failures, rewrite the prompt, and gradually modify the solver itself.

The most important conclusion from this scale is that execution loop and evolution loop are not the same thing. In the first case, an agent simply makes new attempts within the given rules. In the second, the rules themselves, artifacts, or even the structure of the agent change. Therefore, the same word Ralph hides a completely different degree of autonomy, cost, and danger. In practice, this also changes the level of trust in the result.

Where the main risk lies

The main criticism in the article boils down to three things. First, when an external loop only looks external, but the true oracle remains inside the model. Then the agent itself decides that the task is closed and easily exits prematurely. Second, criterion drift: if an agent is allowed to rewrite acceptance criteria, a plan, or the validation layer, it can imperceptibly adjust the task to a convenient solution for itself. Third, accumulation of garbage context, when all development happens in one long session and the quality of reasoning falls.

"Which exactly Ralph?"

The author suggests asking this question first. Before running a Ralph-like architecture, you should determine who declares success, where the criteria are physically fixed, what exactly mutates between iterations, and whether the system has cheap machine-checkable feedback. This is why the author considers the most practical compromise schemes where knowledge and working artifacts can accumulate, but the external verifier and success criteria remain separate and as rigid as possible. Otherwise, the term masks too different engineering solutions.

What this means

For teams building agentic systems, the article is useful as a checklist against confusion. Ralph loop can no longer be used as a universal label: you first need to decide whether you're building an execution loop for reliable execution or an evolution loop with controlled mutation, and only then choose the architecture.