Ralph loop

The Ralph loop is a simple methodology: keep an AI coding agent running in a loop so it can make steady progress over a long sequence of tasks. Instead of one-off prompts, the loop keeps execution focused, repeatable, and aligned with a defined plan.

The concept originates from Geoffrey Huntley, who described it as a simple shell loop that repeatedly runs an agent prompt. Ralph productizes that idea into a full workflow with task tracking, verification, and session memory. Read the original article: Ralph Wiggum as a “software engineer”.

The loop in practice

At its core, the loop is about running the agent repeatedly:

while :; do cat PROMPT.md | claude-code ; done

Ralph turns that pattern into a structured workflow by binding each iteration to a specific task and adding safety rails.

How Ralph applies the loop

Ralph implements the loop with the following structure:

  • A PRD defines the ordered list of tasks
  • Each iteration runs the agent against the next pending task
  • Progress and completion are tracked after every iteration
  • Verification and retries handle failures without losing context
  • Session state persists so work can resume later

Why it matters

Running an agent in a loop provides momentum. Ralph adds the structure needed to keep that momentum aligned with real deliverables, even across long-running sessions.

Next steps