Geoffrey Huntley

Geoffrey Huntley created the Ralph Wiggum Loop, an approach to autonomous coding that repeatedly sends an agent back to a defined task until its work satisfies meaningful checks. His central concern is how engineers turn unpredictable models into dependable systems through precise specifications, durable state, verification, and security boundaries.

Huntley worked across infrastructure, consulting, mobile development, and engineering management before specializing in developer tools and open source. His earlier roles included Telstra, South32, Ansarada, Uno Platform, Coder, and Gitpod; he also co-founded WeeklyXamarin and helped maintain ReactiveUI, the .NET application framework. At Gitpod, he advocated for companies to identify and support the open-source maintainers behind their software dependencies.

He later held technology-lead roles at Optiver and Canva. While leading developer-productivity work at Canva, Huntley recognized that engineers’ repetitive conversations with coding assistants could themselves be automated. He subsequently worked on Sourcegraph’s Amp coding agent and now focuses on Latent Patterns, an educational initiative teaching AI systems and agent fundamentals.

• Ralph Wiggum Loop: Huntley’s original Ralph description reduces agent orchestration to repeated execution against a clear task, with progress preserved in files or version control. Fresh sessions limit accumulated context degradation; explicit stopping conditions and human judgment prevent iteration from becoming aimless.

• Verification as engineering backpressure: Tests, pre-commit hooks, static analysis, compilation errors, and architectural constraints keep agents from advancing until their output meets defined requirements. Huntley regards static types as verification, while recognizing that stronger checks can increase build and testing costs.

• Context-window discipline: Larger context windows do not guarantee useful working memory. Huntley favors tightly scoped prompts, fresh sessions, and external state because long conversations degrade and repeated summarization loses information.

• CURSED and practical agent education: His experimental CURSED programming language tests whether agents can construct compilers, language specifications, and supporting libraries. His coding-agent workshop and open-source implementation teach developers to assemble agents from inference, tools, file operations, and execution loops.

Huntley also emphasizes isolation and software supply-chain control: blocked agents may search for more powerful credentials, and unattended execution can magnify existing security weaknesses. A public self-healing Ralph experiment reflects his interest in agents that detect and repair system failures. At the AI Engineer World’s Fair debate on coding-agent loops, he defended practical automation while rejecting the claim that fully autonomous software factories are already solved. For Huntley, engineering increasingly means specifying desired behavior, designing verification, constraining failure, and remaining accountable for the result.