LOOPS

Self-pacing agent workflows. Copy a loop, paste it into your agent, let it run.

Compatible tools

Works with your agent

Claude Code

Cursor

Codex

Gemini CLI

What are loops

Autonomous agent workflows that iterate until done

A loop is a structured prompt that tells an AI coding agent (Claude Code, Cursor, Codex, Gemini CLI) to repeat a task autonomously until a measurable exit condition is met. Instead of prompting, inspecting, re-prompting manually — you define the goal, the check command, and the exit condition, then let the agent self-pace.

How it works

The loop execution cycle

1

Define goal + exit condition

What does 'done' look like? A passing test suite, a clean audit, a merged PR.

2

Set the check command

The command the agent runs between iterations to evaluate progress.

3

Agent executes a step

The agent does one unit of work — fix a bug, write a test, tighten a constraint.

4

Agent self-evaluates

Runs the check command, reads output, decides: continue or stop.

5

Loop or exit

If exit condition is not met and iterations remain, the agent continues. Otherwise it stops and reports.

5 loops

Build Until Green

Hardened

Run production build, fix compile/bundle errors, repeat until exit 0.

BuildingManual|Claude CodeCursorCodex

Context-Window Packer

Compress context into minimal tokens before handing to an agent. Loop until agent succeeds on first pass.

BuildingManual|Claude CodeCursor

Harness Tightener

After agent run, review for drift from spec. Tighten constraints and re-run until output matches intent.

BuildingManual|Claude CodeCursor

Ship PR Until Green

Hardened

Implement, test, push, open PR, fix CI until all checks pass.

BuildingManual|Claude CodeCursorCodex

Type-Safe Refactor

Rename/restructure with TypeScript as guardrail. Loop until tsc exits 0 with no any escape hatches.

BuildingManual|Claude CodeCursor