Proofloop

Proofloop

  Agents that run until done

  Define what "done" means. Go to sleep. Wake up to verified results.

Why Proofloop

Proofloop is an AI agent orchestrator that solves the main problem with coding agents on real-world tasks: long tasks require constant user-in-the-loop.

When a task takes more than an hour (often several hours), a regular agent becomes a project you need to babysit:

- "continue", "try differently", "now run tests", "fix the regression"
- Manual verification after every step
- Lost context between sessions and iterations
- Subjective "seems done" instead of proven results

Proofloop changes the paradigm: describe what you want (detailed or brief — your choice). During planning, the agent proposes assumptions, risks, a step-by-step plan, and verifiable conditions for "done". You review and approve — then the agent works autonomously until verified completion.

The Problem

Agents alone

Proofloop

Quickstart

1. Install

2. Setup provider (choose one)

Proofloop orchestrates existing AI agents — install whichever you prefer:

OpenCode

Codex (ChatGPT Plus/Pro)

Claude Code

3. Run

Where  is one of: opencode, codex, claude

CLI

Features

- Completion conditions — Automated (pytest, mypy, make build) or text-based ("API returns 200 for all endpoints")
- No limits — Runs for hours, handles 50+ iterations, retries failures automatically
- Fire and forget — Start before bed, wake up to verified results
- Independent verification — Conditions checked by running actual commands, not agent self-assessment
- Smart supervisor — Detects loops, stagnation, regressions; decides retry vs rollback vs stop
- Multi-provider — Uses OpenCode, Codex, or Claude Code under the hood
- Multi-repo — Coordinates changes across multiple repositories

Usage…