Ralph Loops: Build Dumb AI Loops That Ship — Chris Parsons, Cherrypick

May 4, 20261:48:2632,177 views

Summary

Chris Parsons, CTO and consultant at Cherrypick, leads a two-hour hands-on workshop at AI Engineer Europe 2026 on "Ralph Loops" — a deliberately simple pattern for agentic AI automation that prioritizes reliable shipping over architectural sophistication. Parsons argues that most teams gravitate toward multi-agent orchestration, planning graphs, and elaborate tool chains, then spend months debugging complexity rather than delivering value. His central thesis: a dumb loop that processes one ticket at a time, evaluates its own output, and iterates will outperform these systems in practice.

The workshop is structured around live coding using Claude Code, building a command-line Pomodoro timer from a flat-file ticket backlog. Parsons demonstrates the simplest Ralph Loop: a shell-level `while true` loop that invokes Claude Code on each iteration with an instruction like "implement the next most important ticket using TDD principles from doc/tickets, commit when done." Claude reads the ticket list, picks the top unfinished item, writes a failing test, implements the feature, marks the ticket done, and commits — entirely autonomously. Parsons shows this running through five tickets live, with each session starting fresh to avoid context pollution.

A key practical insight is Claude Code's built-in `/loop` command (internally backed by a CronCreate tool), which can schedule an instruction to fire every N minutes without a shell script wrapper. Parsons demonstrates: `loop every 1 hour: check Linear for new bug reports from test`. The cron fires, Claude works, finishes, checks the cron again, and continues indefinitely. He notes this pattern mirrors the entire software development lifecycle — pick ticket, check architecture, make change, submit PR, wait for review, merge, repeat — and argues that most knowledge work is structurally just loops.

Parsons walks through his production Ralph skill file, which encodes: role definition ("you are one energy unit in a relay team, do exactly one change, then drop context"), ticket status values, Git state recovery logic for mid-flight crashes (dirty tree + passing tests = probably done; failing tests = throw away and retry), TDD enforcement, and parallel verification steps. He emphasizes that the skill file is a living document tuned over time for each project — start from his template, then say "fix this for my project" and let it evolve.

On synthetic feedback loops, Parsons describes building a simulate-audience skill that generates multiple client personas from survey transcripts and website responses, runs them as parallel sub-agents to critique newsletter content, and surfaces improvement suggestions without human taste judgments — useful when the "good or bad" question can be delegated to the AI itself.

Sandboxing is discussed in practical terms: he runs most long-horizon loops on a VPS with separate API keys, fine-grained Claude permissions (read-only for most tools, draft-only for email, no send), and is building a project called "lockbox" to block file system access after untrusted token ingestion. Docker sandbox (new container isolation feature) is mentioned as an alternative for local use.

In Q&A, Parsons addresses cognitive debt — the risk of losing track of a codebase the AI has modified — by requiring loops to stop and report before closing any ticket, leaving a human verification gate. He reviews sessions nightly using a Claude skill that saves all daily session logs to JSON, summarizes what was done, and generates improvement suggestions for the loop system itself.

The workshop closes with a discussion of team scaling, theory of constraints (citing Goldratt's "The Goal", 1984), and the observation that AI-assisted teams sometimes go slower because they haven't fixed their release or review bottleneck first. Parsons also mentions an ambitious "startup" skill he is building — a loop designed to run an entire startup decision cycle, which autonomously generated an investor memo deck mid-run without being asked. He views multi-agent orchestration tools like GasTown and MCP Agent Mail as interesting but solving a problem most teams don't yet have.

Key Takeaways

•Start with the dumbest possible loop — a while-true shell script calling Claude Code with one instruction — before reaching for multi-agent frameworks; complexity rarely pays off early.
•Use Claude Code's built-in /loop command to schedule recurring instructions (e.g., 'loop every 1 hour: check Linear for new bug reports') without writing shell scripts.
•Encode your Ralph Loop logic in a skill file that evolves over time: start from a template, then tell Claude to 'fix this for my project' and let it accumulate project-specific test commands, ticket formats, and recovery states.
•Give the loop a Git state recovery heuristic: dirty tree + passing tests = probably done, verify; failing tests = mid-flight, discard or retry.
•Require the loop to stop and report before closing any ticket, keeping a human verification gate to prevent cognitive debt from accumulating unreviewed changes.
•Build synthetic feedback loops by routing output through a separate sub-agent evaluator (e.g., simulate-audience personas), removing yourself from the judgment step entirely.
•Run long-horizon loops on a VPS with separate API keys, read-only tool permissions, and draft-only email access — never let the loop send email autonomously.
•Apply theory of constraints before adding AI tooling: if your bottleneck is the release process or review queue, faster coding loops will make things worse, not better.
•Use nightly session-log summarization (save all Claude Code sessions to JSON, run a skill to summarize and suggest loop improvements) to avoid losing track of autonomous work.
•CICD, linting, and end-to-end tests (Playwright) are force-multipliers for AI loops — they give the agent objective pass/fail signals and reduce the surface area requiring human review.

Transcript

Welcome. So, this workshop is on Ralph loops. Hands up here who knows what a Ralph loop is? That's almost everyone. I'm guessing that the other folks who came in were just here because they thought that sounded weird or maybe looking for a quiet place to work, I don't know. But I'm you're very welcome. So, what we're going to do today, this is a 2-hour workshop. We're going to if you could just kind of make a little bit of space if you need to as people are coming in, that'd be really helpful. T...