Ralph Wiggum loop

A feature regarding Geoff Huntley, his invention of the Ralph Wiggum Loop and the state of coding and agile:

Developers, he argues, should now spend more time thinking about writing loops that drive coding assistants to produce better output, rather than persisting with code reviews.

#!/bin/bash
while :; do cat PROMPT.md | claude-code ; done

Agile and standups doesn't make sense any more," Huntley said. The days of being a Jira ticket monkey are over.

It is not about GitHub Copilot as it used to be year or two ago, it's about Claude Code, OpenAI Codex and Gemini CLI as they started to be from the 2nd half of 2025 on.

It is not about LLM based code completion, it's about AI assistants running automously in your (local) code repository (better in a container) until the task is done.

It is not about just coding by assistants, but also about creating all tests, verifying results based on tests and checking the match to the software architecture and business logic specified in the original prompt.

I little more sophisticated Palrh Wiggum Loop can be found in a posting at anthropic.com where an Anthropic researcher describes how he builda C compiler with a team of parallel Claudes:

#!/bin/bash

while true; do
    COMMIT=$(git rev-parse --short=6 HEAD)
    LOGFILE="agent_logs/agent_${COMMIT}.log"

    claude --dangerously-skip-permissions \
           -p "$(cat AGENT_PROMPT.md)" \
           --model claude-opus-X-Y &> "$LOGFILE"
done

A post at arstechnica.com explains that the $20,000 experiment compiled a Linux kernel but needed deep human management. The posting gives a good overview of the current limitations of that approach and the importance of existing comprehensive test suites or the extensive creation of a full set of tests, that verify the correctness of the resulting code.

Related posts