Releases·Tool Release
Loop engineering: turning the Ralph loop into a verified loop with LOOP.md, a critic veto, and protected files
A developer shared "loop," a tool that upgrades the Ralph loop — the while-loop that repeatedly pipes PROMPT.md into Claude — into a verified version. The original loop lacks memory between runs, trusts the agent's claim of "done," never stops itself, and allows the agent to edit its own tests. The tool adds a critic step before completion, framed by the author as "loop engineering": controlling what happens after the agent answers rather than the prompt itself.
3 engagement1 sourceMon, Sep 7, 2026, 11:49 PM
AI Overview
- loop (github.com/raiyanyahya/loop) keeps goal, done criteria, protected files, critic, and metric in one LOOP.md frontmatter
- Works with Claude Code, Codex, Gemini CLI, Aider, or any CLI agent; zero dependencies, MIT-licensed
- Agent writes a letter to its next self between iterations as the only memory across amnesiac runs
- Claude Haiku run: three iterations, premature "done" rejected on open checklist items, critic approved diff, total 2m13s
- LOOP.md options include until, protect (e.g. "test/**"), independent critic (e.g. codex), and a bench metric that keeps or reverts iterations
Entities
Related
Social·Community
Loop Engineering: designing systems that prompt coding agents instead of hand-prompting them
A Substack essay, shared on Reddit, advocates 'loop engineering': building automated systems that prompt coding agents recursively until a verified goal is met, rather than hand-writing prompts. It catalogs five building blocks now shipping in Claude Code and Codex, plus on-disk memory as a sixth. For practitioners, it reframes agent work as system design around parallelism, skills, and verification loops.
A Substack essay, shared on Reddit, advocates 'loop engineering': building automated systems that prompt coding agents recursively until a verified goal is met, rather than hand-writing prompts. It catalogs five building blocks now shipping in Claude Code and Codex, plus on-disk memory as a sixth. For practitioners, it reframes agent work as system design around parallelism, skills, and verification loops.
2 engagement1 sourceredditFri, Sep 4, 2026, 07:00 AM
Products·launched
Loop-native canvas for Claude Code to run in loops with other agents
Looperators is an open-source canvas that automates multi-agent loops. When Claude Code finishes a task, its diff is automatically sent to a reviewer agent; blocking issues wake the same Claude session to fix them, and the loop stops after a configurable number of laps. It solves the problem of manually orchestrating iterative code review and revision workflows for developers using Claude Code.
Looperators is an open-source canvas that automates multi-agent loops. When Claude Code finishes a task, its diff is automatically sent to a reviewer agent; blocking issues wake the same Claude session to fix them, and the loop stops after a configurable number of laps. It solves the problem of manually orchestrating iterative code review and revision workflows for developers using Claude Code.
3 engagement1 sourceredditTue, Jul 21, 2026, 06:44 PM
Products·building
Extended Gauntlet Loop for Production Apps and CI
This project extends Matt Shumer's Gauntlet Loop (Claude of Duty) to handle production-ready applications, integrating with CI, migrations, and a protected production branch. It uses LLM agents as builders and critics within each task to ensure quality against a reference, aiming to make the loop more practical for real-world development.
This project extends Matt Shumer's Gauntlet Loop (Claude of Duty) to handle production-ready applications, integrating with CI, migrations, and a protected production branch. It uses LLM agents as builders and critics within each task to ensure quality against a reference, aiming to make the loop more practical for real-world development.
1 engagement1 sourceredditWed, Aug 19, 2026, 04:18 PM
Products·launched
Deterministic checker for Claude Code agent claims vs reality
Agent Vigil is a deterministic tool that verifies whether Claude Code agents actually did what they claimed in their session logs. It reruns tests, diffs file changes, checks paths, and cross-references tool calls, without using LLM judges. It helps developers trust agent outputs by catching discrepancies.
Agent Vigil is a deterministic tool that verifies whether Claude Code agents actually did what they claimed in their session logs. It reruns tests, diffs file changes, checks paths, and cross-references tool calls, without using LLM judges. It helps developers trust agent outputs by catching discrepancies.
5 engagement1 sourceredditThu, Aug 20, 2026, 03:04 PM
Social·Community
Developers share agentic loop patterns beyond Huntley's ralph loop
A developer asks how others use agentic loops, citing Geoffrey Huntley's foundational ralph loop pattern. The discussion surfaces evolved practices and tools, including the /loop slash command, while noting context bloat concerns.
A developer asks how others use agentic loops, citing Geoffrey Huntley's foundational ralph loop pattern. The discussion surfaces evolved practices and tools, including the /loop slash command, while noting context bloat concerns.
4 engagement1 sourceredditMon, Aug 3, 2026, 07:41 PM