Geoffrey Huntley

Geoffrey Huntley, also known as Geoff, is the creator of the Ralph Loop, a technique for repeatedly running a coding agent against specifications and an implementation plan. His work combines autonomous code generation with engineering controls: bounded tasks, validation feedback, and human intervention when the agent builds the wrong thing. He also teaches developers how to build coding agents themselves.[1][5][3][6]

In his Ralph explanation, first published in July 2025, Huntley writes: “Ralph is a technique. In its purest form, Ralph is a Bash loop.” The minimal example repeatedly feeds PROMPT.md to a coding agent. The fuller approach operates in a single repository through one primary process, loading specifications and a plan each iteration. Huntley generally recommends one item per loop, with permission to relax that limit as the project progresses and narrow it again when results deteriorate.[3]

The single primary process can delegate work to subagents. Huntley recommends using its context as a scheduler, moving expensive searches and test-result summaries into delegated work. Specifications and the implementation plan provide continuity between iterations; brief project instructions record discoveries about how to build and run the software. He also asks agents to document why tests matter, so later iterations retain reasoning that would otherwise disappear with the previous context.[3]

Huntley calls the feedback that rejects invalid generated code backpressure. Builds, focused tests, type systems, static analyzers, and security scanners can all supply it. His prescription balances correctness against the speed of the feedback cycle. During his CURSED programming-language experiment, he says he chose Rust because he wanted extreme compiler correctness, accepting slower compilation and more attempts to produce working code.[3]

CURSED also supplied concrete lessons about operator error and recovery. Huntley recalls discovering a month into the work that his lexer specification defined one keyword for two opposing scenarios, wasting time until he recognized the contradiction. He repeatedly discarded and regenerated the implementation plan when it became unhelpful. When compiler errors filled Claude’s context, he says he gave the errors to Gemini and asked it to create a recovery plan for Ralph.[3]

Huntley’s recommendation carries substantial limits. In that account, he favors bootstrapping greenfield projects, estimates that the technique can get them roughly 90% complete, and warns against using it in an existing codebase. He describes broken builds that require an engineer to decide whether to reset the work or rescue it through new prompts. His qualification is explicit: “There is no way this is possible without senior expertise guiding Ralph.”[3]

At AI Engineer World’s Fair 2026, Huntley presented the pro-loop case alongside Keycard CEO Ian Livingstone in a debate moderated by Allie Howe. Dex Horthy of HumanLayer and Greg Pstrucha of Subroutine presented skeptical arguments. Huntley said, “I don’t see myself going back to writing code by hand.” His analogy for the engineer’s role was keeping a locomotive on its rails, while acknowledging that software factories were still “frontier thinking” and not yet solved in the market.[1]

His enthusiasm included a warning about premature automation. As reported by Latent Space, Huntley worried that the following year’s conference would bring people saying their factories and loops had failed. “These are things that we are still yet to figure out,” he said.[2]

Huntley published his free coding-agent workshop in August 2025, accompanied by a Go workshop repository. It starts with a basic Claude chatbot and progressively adds file reading, directory listing, shell execution, editing, and code search. The central mechanism is a model/tool feedback loop: the model requests a tool, the application executes it, and the result returns to the model for further inference. One example has the agent create a FizzBuzz script and execute it to check its behavior.[6][7]

In the workshop introduction, Huntley describes moving from developer-productivity leadership at Canva to engineering work at Sourcegraph building Amp. He explains the team’s Oracle as GPT exposed as a tool that Claude Sonnet can call for guidance, checking, research, and planning. His educational argument is that building an agent helps developers move from consuming AI tools to producing automation. He also advances a stronger personal view: AI-related job loss reflects insufficient personal development and self-investment.[6]

In a separate announcement published July 19, 2025, Huntley announced that he had joined X as a security engineer conducting red-team security research on Grok.[4]

Before his coding-agent work, Huntley worked on developer tools and infrastructure. His résumé describes developing cloud-provider Terraform templates at Coder and serving as Gitpod’s sole site-reliability engineer in the JAPAC region while its primary European timezone was asleep. These earlier responsibilities provide context for his experience with build systems and development environments.[5]

Huntley says he spent six years as lead and core maintainer of ReactiveUI, a cross-platform .NET framework inspired by functional reactive programming. In a testimonial reproduced on Huntley’s site, You, Me and ReactiveUI author Kent Boogaart credits him with automating builds, improving project management, formalizing documentation infrastructure, and helping bring the project into the .NET Foundation.[5]