AI Auto-Work
An agentic development workflow that pairs Claude Code as the executor with Codex as an adversarial reviewer, taking a requirement through research, planning, development and cross-model review with mechanical compile-and-test gates before each atomic commit.
AI Auto-Work is a workflow system that orchestrates large language models to carry a software change through a full cycle: requirement research, research review, planning, plan review, development and development review, aimed at handling medium to large engineering tasks with minimal human intervention.
Claude Code acts as the executor that writes, fixes and tests code, while Codex acts as an adversarial reviewer checking for concurrency issues, resource leaks and edge cases, forming what the project calls a triple-check convergence loop; a mechanical quality gate requires the code to compile and pass tests before any review cycle begins. Work above a size threshold is automatically split into medium-sized tasks, each run in an isolated process so context from one stage does not leak into the next, with handoff only through persisted files.
Each atomic commit is limited to at most three implementation files, eight total paths including tests and manifests, and roughly 100 lines, so history stays bisectable, and systematic errors get written back into a shared knowledge base rather than only fixed in the code.