Use Autonomous Loops for Hands-Off Development

The autonomous loop pattern (popularized as “Ralph Wiggum” in the community) lets Claude work independently in a cycle: build, test, fix, repeat.

The Basic Loop

#!/bin/bash
# autonomous-loop.sh
MAX_ITERATIONS=10
PROMPT="$1"

for i in $(seq 1 $MAX_ITERATIONS); do
  echo "=== Iteration $i/$MAX_ITERATIONS ==="
  claude -p "$PROMPT" --allowedTools "Read,Write,Edit,Bash" 2>&1

  # Circuit breaker: stop if tests pass
  if npm test 2>&1 | tail -1 | grep -q "passed"; then
    echo "All tests passing. Done."
    exit 0
  fi
done

echo "Hit max iterations without success."
exit 1

Usage

chmod +x autonomous-loop.sh
./autonomous-loop.sh "Implement the user settings API endpoint. \
  Write tests first, then implement until all tests pass. \
  Run npm test after each change."

Circuit Breakers

Always include a way to stop the loop:

• Test passing: stop when the test suite is green
• Max iterations: hard cap to prevent runaway costs
• Error threshold: stop after 3 consecutive failures
• Cost cap: check token usage between iterations

Headless Mode Version

claude -p "$PROMPT" \
  --allowedTools "Read,Write,Edit,Bash(npm test*)" \
  --max-turns 20

The --max-turns flag acts as a built-in circuit breaker for headless mode.

When to Use

• Implementing a feature with a clear test suite
• Fixing a batch of linter errors
• Migrating code to a new pattern across many files

When NOT to Use

• Tasks without clear success criteria (no tests, no linter)
• Architectural decisions that need human judgment
• Anything touching production infrastructure