🔁 Loop Library

The most complete, credited collection of AI-agent loops — repeatable workflows you hand to a coding agent (Claude Code, Cursor, Codex, Gemini CLI…) that iterate a process until a stopping condition is met.

Refactor until the architecture is clean. Add tests until 100%. Hunt bugs until none remain. You bring the vision — the loop brings the senior engineer.

🔎 Browse & copy every loop interactively → keysersoose.github.io/loop-library

Every loop ships with a ready-to-paste prompt. Search, filter, copy one — or grab them all: ⬇ Download all 156 prompts (ALL-LOOPS.md) · loops.json (machine-readable) · one click, no link-chasing.

---

What is a loop?

A loop is a prompt that doesn't stop at "done once." It repeats a process until a condition is met. Progress lives in your files, tests, and git history — not in the chat window — so the agent can run fresh context every pass and keep going long after a single prompt would have quit.

A good loop specifies five things (framing credit: Forward Future Loop Library):

Part	Question it answers
Trigger	When does the loop start?
Action	What does it do each pass?
Proof	How does it verify progress — a test, a scan, a screenshot — not its own opinion?
Memory	What carries state between passes (files, a TODO, git)?
Stopping condition	When does it exit — and what's the safety cap so it can't spin forever?

flowchart LR
  T([Trigger]) --> A[Action]
  A --> P{Proof?}
  P -- not yet --> M[(Memory:<br/>files · TODO · git)]
  M --> A
  P -- passed --> X([Exit ✓])
  A -. safety cap .-> S([Stop &amp; ask])

Loading

This repo is a credited index. Every loop links to its original source and names its original creator — the loops belong to the people who made them. Wrong credit? Open an issue and we'll fix it immediately.

---

Contents

• ⭐ Start here — foundational loops
• Engineering
• Testing & evaluation
• Prompt & model optimization
• Research & data science
• Security & red-teaming
• Writing & content
• Design
• Accessibility
• Data & analytics
• International (translated)
• Marketing & growth
• Support & sales
• Operations
• DevOps & infrastructure
• Autonomous coding agents (loop-based)
• Loop frameworks (GitHub)
• 🛠 Tools & harnesses (run-it-yourself loop repos)
• 📚 Patterns & theory (the research behind loops)
• Full loops included in this repo
• Credits & sources
• Contributing
• License

---

⭐ Start here — foundational loops

The patterns that defined loop-driven development.

Loop	What it does	Creator	Source
The Ralph (Ralph Wiggum) Loop	Feed the same prompt to an agent in an infinite shell loop; each pass gets fresh context, state survives in files + git. The origin of loop-driven development.	Geoffrey Huntley	ghuntley.com/loop
autoresearch	Point an agent at a measurable script; it proposes a change, runs a time-boxed (5-min) experiment, keeps it if the metric improves, reverts if not — all night. 66k+ stars.	Andrej Karpathy	explainer · awesome list
The Senior Loops ⭐	Five build loops (spec → plan → test → security → critic) that force AI output to senior-engineer quality. Full text in this repo.	keysersoose	loops/senior-loops.md

---

Engineering

Loop	What it does	Creator	Source
The Docs Sweep	Review the codebase and bring every doc back in sync with the real implementation.	Matthew Berman	Forward Future
The Architecture Satisfaction Loop	Refactor — with live tests and commits each step — until the architecture is genuinely clean.	Peter Steinberger	Forward Future
The Sub-50ms Page-Load Loop	Optimize until every page loads in under 50 ms.	Matthew Berman	Forward Future
The Production Error Sweep	Read production logs, trace each error to its root cause, and fix it.	Matthew Berman	Forward Future
The 100% Test Coverage Loop	Add tests until coverage reaches 100%.	Matthew Berman	Forward Future
The Logging Coverage Loop	Add logging until every important path emits useful, tested logs.	Matthew Berman	Forward Future
The Nightly Changelog Loop	Each night, summarize the previous day's changes into the changelog.	Matthew Berman	Forward Future
The Test-Suite Speed Loop	Make the test suite run as fast as it can.	Matthew Berman	Forward Future
The Repository Cleanup Loop	Audit branches, PRs, commits, and worktrees, then tidy up.	Matthew Berman	Forward Future
The Ticket-to-PR-Ready Loop	Turn a ticket into a review-ready patch with the failure defined up front.	Hiten Shah	Forward Future
The Clodex Adversarial-Review Loop	Run a Codex code review and fix every finding above a set threshold.	Lukas Kucinski	Forward Future
The Loop Harness Verification Loop	Have a second agent session verify the staged work against the criteria.	Istasha	Forward Future
The Fresh-Clone Loop	Clone into a clean, empty environment and follow the README — to catch setup gaps.	0xUmbra	Forward Future
The LLM Codegen Workflow	Brainstorm a spec (LLM asks one question at a time) → generate a prompt_plan.md + todo.md → feed prompts to the agent in order.	Harper Reed (popularized by Simon Willison)	harper.blog
The Multi-File Refactor Loop	Plan the change, apply rule-based transforms across call sites, run tests after each, repeat until behavior is preserved and clean.	(community)	Augment Code
The Code Migration Loop	"Environment-in-the-loop": plan → set up env → migrate → run tests → refine, iterating until the port passes.	(research)	arXiv
The Self-Correcting Agent Loop	A state-machine loop (LangGraph/CrewAI) with conditional edges: act → check → retry, bounded by iteration limits.	(community)	guide
The Mobile Closed-Loop Build	Install on a virtual device, tap through every screen, verify layouts across sizes, fix crashes, repeat until store-ready.	Alexander Zanfir	writeup
Timebox / Ultimatum Loop	Give an assistant's suggestion a hard time budget; if it doesn't yield value with minimal extra effort, abandon and code manually.	Birgitta Böckeler (Thoughtworks)	martinfowler.com
Guardrails Correction Loop	Generate → validate; if a check fails, filter or regenerate; stop when output passes all validators or a max-retry cap.	Eugene Yan	eugeneyan.com

---

Testing & evaluation

Loop	What it does	Creator	Source
The Quality Streak Loop	Test realistic scenarios; when one fails, document it and fix it.	Matthew Berman	Forward Future
The Full Product Evaluation Loop	Generate N realistic scenarios, each with explicit success criteria.	Matthew Berman	Forward Future
The Self-Improving Champion Loop	Iterative optimization with a budget, scoring, and gate checks.	Jose C. Munoz	Forward Future
The Devil's-Advocate Loop	Argue against your own design — via a critic sub-agent — until it survives.	Anonymous	Forward Future
TDAD — Test-Driven Agentic Development	Use a code dependency graph for blast-radius before each change; cut agent regressions.	(research)	arXiv
Eval-Driven Development (LLM-as-Judge)	Write evals first → change → run evals → integrate; route flagged production traces back into a golden dataset.	Eugene Yan (& community)	eugeneyan.com
Pre-Commit Guard	A hook that runs the test suite before every commit and blocks the commit while the suite is red.	elorm	loops.elorm.xyz
Post-Edit Test Guard	A hook that runs the related tests after each file edit to catch regressions immediately.	elorm	loops.elorm.xyz
Independent Verifier Pass	When the implementer claims done, a separate verifier runs build + lint + tests with no access to the implementer's rationale.	elorm	loops.elorm.xyz
Plan-Execute-Validate (PEV)	Planner → executor → LLM-scorer; route to retry / replan / complete by quality score; stop on pass or budget.	Manjunath G	dev.to
The Iteration Flywheel	Evaluate quality → debug → change behaviour; tighten the eval step to speed every other step.	Hamel Husain	hamel.dev
Model–Human Alignment Loop	Iterate the LLM-judge critique prompt until model grades agree with human grades on 25–50 examples.	Hamel Husain	hamel.dev
TDD Red-Green-Refactor (AI-assisted)	Red→Green→Refactor for AI coding; each phase gates on a hard condition; fall back to manual if regen fails twice.	Paul Sobocinski (Thoughtworks)	martinfowler.com

---

Prompt & model optimization

Loops that improve the prompt or model itself against a metric.

Loop	What it does	Creator	Source
OPRO (Optimization by PROmpting)	Feed the LLM prompts paired with their scores; it generates better variants, iterating toward higher accuracy.	Google DeepMind (Yang et al.)	overview
TextGrad	"Textual gradient descent" — an LLM proposes edits that reduce a textual loss, backpropagated through your pipeline.	Stanford (Yuksekgonul et al.)	arXiv
DSPy	Program prompts as modules with signatures; a compiler optimizes them against a metric instead of you hand-tuning.	Stanford NLP (Khattab et al.)	github
PromptBreeder	Evolutionary search that mutates and "breeds" prompts, selecting by fitness across generations.	Google DeepMind (Fernando et al.)	overview

---

Research & data science

Loop	What it does	Creator	Source
autoresearch	The headline ML loop: edit → time-boxed experiment → measure → keep/revert → repeat.	Andrej Karpathy	awesome-autoresearch
The AI Scientist	Autonomous loop: generate research ideas, run experiments, analyze results, write the paper.	Sakana AI (Lu et al.)	survey
Agentomics-ML	Autonomous ML experimentation loop for genomic/transcriptomic data.	(research)	arXiv
Deep Research Loop	Planner → search sub-agents → synthesize → "need more?" → loop until coverage is sufficient, then cite.	LangChain (Open Deep Research)	blog
The MLOps Retraining Loop	Monitor for drift → when it exceeds threshold, retrain → compare vs current → ship only if better → keep monitoring.	(community)	guide
DeepSearch Loop	Iterative search → read → reason over vector/FTS/ripgrep tools; loop until the optimal answer is found.	Han Xiao (Jina AI)	via Simon Willison

---

Security & red-teaming

Loop	What it does	Creator	Source
The AI Red-Teaming Loop	Adversarially attack the system (prompt injection, jailbreaks, data leakage), feed results back, re-test on a schedule/CI.	(community)	AI-Red-Teaming-Guide
RvB — Iterative Red-Blue Games	Automate system hardening through repeated red-team-attacks / blue-team-defends rounds.	(research)	arXiv

---

Writing & content

Loop	What it does	Creator	Source
Self-Refine	One LLM drafts, critiques its own draft, then revises — looping until a stopping criterion. ~20% avg gains across tasks.	Madaan et al.	learnprompting
Draft → Critique → Finalize	Force deliberation: generate an idea, critique it against criteria, regenerate from the critique.	(community)	learnwithparam
LLM Peer Review	Multiple agents review each other's drafts, then privately revise — distributed critique for creative writing.	(research)	arXiv
Chain of Density	Iteratively rewrite a summary, fusing in missing entities each pass without growing the length — denser every loop.	Adams et al.	arXiv · learnprompting
The SEO/GEO Visibility Loop	Audit crawlability, indexation, and structured data.	Matthew Berman	Forward Future
The Product Update Podcast Loop	Generate a short 3–5 minute podcast episode about new features.	Pierson Marks	Forward Future

---

Design

Loop	What it does	Creator	Source
War Loops: Autonomous Frontend Designer	Point it at a URL or image and produce iterative design builds.	Swayam	Forward Future
The Boeing 747 Benchmark	Build the most realistic Boeing 747 you can in Three.js — a model-capability benchmark.	@victormustar	Forward Future
The Infinite Clickbait Loop	Generate ten thumbnail concepts and score each against a rubric.	@Alex_FF	Forward Future
The Three.js Game Loop	Build a playable loop first, then iterate (playtest → tweak → re-test) until it passes gameplay, visual, perf & release checks.	majidmanzarpour	github
VideoAgent	Agentic video understanding/editing with reflection rounds + adaptive feedback loops that refine the plan via self-evaluation.	HKUDS	github

---

Accessibility

Loop	What it does	Creator	Source
The a11y Audit Loop	Scan for WCAG violations, apply remediation hints, re-scan until compliant. (Automated tools catch ~30–40% — pair with manual testing.)	snapsynapse	skill-a11y-audit
Accessibility Review Agents	Specialist agents that enforce WCAG 2.2 AA so coding agents stop generating inaccessible code.	Community-Access	github

---

Data & analytics

Loop	What it does	Creator	Source
The Text-to-SQL Self-Correction Loop	Creator generates SQL → Runner executes → Enhancer critiques → regenerate, until the query is executable and semantically correct.	Rudder Analytics	writeup
The Active-Learning Labeling Loop	Auto-label high-confidence items, route low-confidence ones to humans, retrain, repeat until a quality/budget stop rule is hit.	(community)	guide

---

International (translated)

Loops sourced from Chinese-language communities (53AI, Tencent Cloud), translated to English with credit to the original authors.

Loop	What it does	Creator	Source
Loop Engineering (循环工程)	Write the program that drives the agent: goal → execute → judge → re-prompt with errors → repeat; stop on budget or judge pass.	Omar (omarsar0); ThinkInAI	53AI
MetaSkills Self-Bootstrap Loop (元技能自举)	Agent generates new reusable workflow templates for itself: match → fill → DAG check → quality gate → sandbox → human approval.	geffzhang; via 53AI	53AI
Judge-Model Termination Loop (法官模型)	Pursue a persistent goal across turns with a separate judge model deciding completion (not self-declaration).	Youming Zhang, Tencent Cloud	Tencent Cloud
Self-Repair Loop (pass@t)	Generator → executor (tests) → separate feedback model → repair model; stop when a sample passes all tests.	Gao & Wang, Microsoft Research	writeup
Prefix-Caching Agent Loop (前缀缓存)	Tool-call loop where the old prompt stays an exact prefix of the new one, making per-token cost linear via KV-cache reuse.	OpenAI Eng; via Synced	writeup

---

Marketing & growth

Loop	What it does	Creator	Source
Loop Marketing / Growth Loop	Run small, fast A/B tests on hooks, offers and audiences; roll each learning back into creative, targeting and spend — compounding every cycle.	HubSpot	hubspot.com

---

Support & sales

Loop	What it does	Creator	Source
The Support Resolution Loop (ATTD)	Analyze → Train → Test → Deploy: resolve tickets, learn from human-corrected escalations, and improve resolution over time.	Fin AI (Intercom)	fin.ai

---

Operations

Loop	What it does	Creator	Source
The Stale-Safe Batch Release Loop	Exclude stale/unfinished work, combine the valid changes, and release.	Matthew Berman	Forward Future
The Production Data Cleanup Loop	Remove anything that doesn't meet the allowed definition.	Matthew Berman	Forward Future
The Post-Release Baseline Loop	Run the standard benchmarks and record the results.	Matthew Berman	Forward Future
The Customer AI Deployment Loop	Manage one customer AI deployment from priority to production.	AgentLed.ai	Forward Future

---

DevOps & infrastructure

Loop	What it does	Creator	Source
The Terraform Fix Loop	validate → plan → apply → verify — read the error, fix, re-run until the infra is correct. (Constrain scope: fix mode loves to refactor adjacent code.)	(community)	guide
Ship PR Until Green	Implement on a branch, run tests, push, open a PR, wait for CI, and loop until checks pass and it's merge-ready.	elorm	loops.elorm.xyz
CI Failure Watcher	Poll CI on an interval; when checks go red, investigate and push fixes until green.	elorm	loops.elorm.xyz
Deploy Verification Loop	After a deploy, hit health + smoke endpoints on an interval until everything returns healthy.	elorm	loops.elorm.xyz

---

Autonomous coding agents (loop-based)

The agent frameworks whose whole architecture is a loop.

Agent	The loop it runs	Creator	Source
Aider	Pair-programming loop with an architect → editor two-model mode; edits via diffs, runs against a tree-sitter repo map.	Paul Gauthier	github
SWE-agent	Agent-computer-interface loop that reads an issue, edits the repo, runs tests, and iterates to a patch.	Princeton NLP	github
OpenHands (ex-OpenDevin)	Multi-agent delegation loop: plan → act → observe across a dev environment.	All-Hands-AI	github
AutoGPT	The original autonomous goal loop: decompose a goal into tasks and execute them until the objective is met.	Significant-Gravitas	github
BabyAGI	Task-creation/prioritization/execution loop backed by a vector store of prior results.	Yohei Nakajima	github

---

Loop frameworks (GitHub)

Battle-tested open-source repos whose whole job is to run a loop for you — verified live, with star counts.

Repo	The loop it runs	Creator	Stars	Source
claude-task-master	Parse a PRD into a dependency-ordered task graph; next returns the unblocked task; implement → done → next until the graph is drained.	eyaltoledano	~27.6k	github
feature-dev (Claude Code plugin)	7-phase gate loop: discovery → explore → clarify → architecture → implement → review → summary, with human gates before architecture & implement.	Anthropic	~133k	github
agent-skills	Slash-command suite (/spec /plan /build /test /review /ship) — a verifiable multi-phase loop; /build auto removes human stepping.	Addy Osmani	~64k	github
claude-engineer	v3 self-expanding loop: detect a capability gap → generate a new tool → validate → hot-reload → chain tools; bounded by a token cap.	Doriandarko	~11k	github
ccg-workflow	Classify → parallel Codex+Gemini analysis → plan → hard-stop user approval → parallel impl → dual-model cross-review → gates.	fengshao1227	~5.6k	github
ouroboros	Socratic interview → immutable spec → decompose → execute → evaluator scores → loop until spec satisfied.	Q00	~4.6k	github
EvoAgentX	Self-evolving agents: evaluators score, then TextGrad/AFlow/EvoPrompt evolve prompts + workflow across iterations.	EvoAgentX	~3.1k	github
agentic-context-engine (ACE)	Run → Reflector analyzes the trace → SkillManager updates a persistent Skillbook → restart with it injected; stops when tests pass or after N epochs.	kayba-ai	~2.5k	github
autoharness	An outer agent improves your harness (prompts, params, context), benchmarks each change on a fixed eval, keeps only wins — overnight.	kayba-ai	~295	github
lauren	Drains a mutable task queue; each task runs implement→review→fix in its own git worktree and auto-merges; edit the queue mid-run.	ofux	~9	github
claude-code-harness	Enforces Plan → Work → Review → Ship; spec approved before code, review gated, evidence packaged before release.	Chachamaru127	~2.8k	github
awesome-harness-engineering	Reference index of agent-loop primitives: planning, context, MCP, memory, orchestration, verification, HITL.	ai-boost	~1.9k	github
continuous-claude	Runs Claude Code in a loop: opens PRs, waits for CI, merges, next task — context via a markdown handoff file.	AnandChowdhary	~1.4k	github
cavekit	grill → spec → research → review → build over a durable SPEC.md; test failures back-propagate into spec invariants.	JuliusBrussee	~1k	github
loki-mode	Multi-agent autonomous SDLC: Reason-Act-Reflect-Verify cycles + quality gates, from PRD to deployed app.	asklokesh	~982	github
ralph-loop-agent	Wraps the Vercel AI SDK in a Ralph outer loop: run → verifyCompletion → feedback → repeat until done/cap.	vercel-labs	~800	github
helixent	Minimal TypeScript/Bun library for ReAct-style think→act→observe loops with parallel tool calls.	MagicCube	~582	github
bernstein	Manager decomposes goal → agents in isolated worktrees → janitor verifies (tests/lint/types) → merge; loop until a predicate passes.	Alex Chernysh	~574	github
self_improving_coding_agent	Evaluate on benchmarks → archive → let the agent improve its OWN codebase → repeat, stacking gains.	MaximeRobeyns	~353	github
orch	CTO agent splits goal into tasks; parallel engineering agents in isolated worktrees; QA verifies; state machine to done.	oxgeneral	~83	github
foreman	Boris-style TUI supervising headless Claude Code agents: plan → ADR/PRD → issues → TDD → e2e, human gates + budget caps.	VisionForge-OU	~56	github
millrace	Governed agentic-loop runtime: stage gates, durable queues, recovery rules that persist state across context windows.	tim-osterhus	~48	github
great_cto	Spec synthesis → single human approval gate → scaffold/backend/frontend/test/deploy; risk-tiered gates; stop when a live URL ships.	avelikiy	~41	github
forge	Idea → R-numbered spec → task DAG → TDD per worktree → reviewer + 4-level verifier; stop on FORGE_COMPLETE or budget.	LucasDuys	~29	github
llm-loop	Plugin for the llm CLI: turn-based autonomous loop (--max-turns, default 25) with optional tool-call approval.	nibzard	~17	github
pi-ralph	Agent cycles through named "hats" (roles) triggered by emitted events; stop on a completion-promise string or guard rails.	samfoy	~13	github

---

🛠 Tools & harnesses (run-it-yourself loop repos)

Open-source repos that run the loop for you.

Repo	What it gives you	Creator	Source
loop-engineering	Patterns, starters & CLI tools (loop-audit, loop-init, loop-cost) for loop engineering.	Cobus Greyling	github
autoloop	Agent-agnostic iterative-optimization loops for Claude Code, Codex, Cursor, OpenCode, Gemini CLI.	Arm Gabrielyan	github
agentic-loop	Autonomous Claude Code toolkit based on RALPH + PRD-driven development.	allierays	github
agent-loop	A guide + workflows for orchestrating specialized AI personas to build, test, and deploy in a loop.	Saik0s	github
ralph	An autonomous loop that runs until every item in a PRD is complete.	snarktank	github
ralph-claude-code	A Ralph loop for Claude Code with intelligent exit detection.	Frank Bria	github
codex-autoresearch-harness	Bash harness to run Karpathy's autoresearch with Codex CLI + an A/B framework.	SarahXC	github
autoresearch-guide	3,000+ line guide teaching agents to autonomously optimize anything measurable.	Rkcr7	github
awesome-autoresearch	A curated list of autoresearch resources.	yibie	github

---

📚 Patterns & theory (the research behind loops)

The academic patterns every loop is built on.

Pattern	The idea	Origin	Source
ReAct	Interleave Thought → Action → Observation so reasoning guides tool use and vice versa.	Yao et al.	explainer
Reflexion	Actor + Evaluator + Self-Reflection; verbal self-feedback stored in memory across trials. 91% pass@1 on HumanEval.	Shinn et al.	NeurIPS
Self-Refine	A single LLM as generator, critic, and refiner, looping with feedback memory.	Madaan et al.	learnprompting
Evaluator-Optimizer	Generator proposes, evaluator scores, loop until a quality threshold — the generic critic loop.	(Anthropic, Building Effective Agents)	pattern catalogue
Self-Consistency / Tree of Thoughts	Sample many reasoning paths / search a tree of thoughts, then select the best.	Wang et al. / Yao et al.	pattern catalogue
Agent Design Pattern Catalogue	A catalogue of architectural patterns for foundation-model agents.	Liu et al.	arXiv
Anthropic's Agent Workflow Patterns	The five canonical patterns: prompt chaining, routing, parallelization, orchestrator-worker, and evaluator-optimizer.	Anthropic	anthropic.com
Basic Reflection	Generator LLM paired with a reflector LLM playing "teacher"; generate → reflect → revise; stop at an iteration cap.	Ankush Gola / LangChain	langchain
Language Agent Tree Search (LATS)	Reflection + Monte-Carlo tree search: Select → Expand → Reflect/Evaluate → Backpropagate; stop when solved or depth cap.	Zhou et al. / LangChain	langchain
On-the-Loop (Harness Loop)	Humans improve the harness that produces artifacts rather than fixing artifacts directly; agents run the inner loop.	Kief Morris (Thoughtworks)	martinfowler.com

---

Full loops included in this repo

Loops whose complete, copy-paste text lives here (authored by this repo, MIT-licensed):

• Senior Loops — the five build loops with exact enter / continue / exit / safety conditions and per-step verification artifacts. Drop into your CLAUDE.md and your AI builds like a senior engineer.

Want the full text of your loop hosted here? See CONTRIBUTING.md.

---

Credits & sources

This library stands on the shoulders of the people who invented and shared these loops. Deep thanks to:

• Forward Future Loop Library — source for many loops here and the trigger / action / proof / memory / stopping-condition framing. Contributors there include Matthew Berman, Peter Steinberger, Hiten Shah, Pierson Marks, Lukas Kucinski, Istasha, 0xUmbra, Jose C. Munoz, AgentLed.ai, @victormustar, Swayam, and @Alex_FF.
• Geoffrey Huntley — the Ralph loop & loop-driven development.
• elorm — the loops! library (CI / testing / deploy loops).
• Andrej Karpathy — autoresearch.
• The researchers behind ReAct, Reflexion, Self-Refine, OPRO, TextGrad, DSPy, and PromptBreeder, and the maintainers of every tool linked above.

Attribution wrong or want your loop linked differently (or removed)? Open an issue — we fix these first.

---

Contributing

Found a loop? Made one? PRs and issues welcome — see CONTRIBUTING.md. One loop per row: name, one-line description, creator, source link. Star the repo if it saved you time — it's how other builders find it. ⭐

License

The curation, descriptions, and original loops in this repo (e.g. loops/senior-loops.md) are MIT licensed. Linked loops belong to their original authors under whatever terms their sources set — this repo indexes and credits them, it does not relicense them.