# Humanize workflow adds layered review to AI code generation

> The new Humanize system uses separate builder and reviewer agents with human‑approved plans, aiming to catch bugs that single agents miss.

Oossa · 2026-10-08 · https://oossa.com/en/humanize-workflow-adds-layered-review-to-ai-code-generation

Researchers led by Sihao Liu released Humanize, a multi‑agent workflow for AI‑assisted coding. A human signs off a plan, a builder agent writes code, and a reviewer agent from a different vendor decides when the work is complete. The system enforces 72 mechanical checkpoints and alternates between the two agents, so a defect survives only if both miss it. In early deployment it collected 1,468 GitHub stars and handled a 567‑file gem5 migration under upstream review.

## The facts

- Published on 2026-10-08 ("Tue, 6 Oct 2026 17:54:42 UTC" shows submission; the paper lists "Thu Oct 08 2026" as publication date).
- Gathered 1,468 GitHub stars after 68 versions in 108 days.

## Why it matters

It offers developers a way to reduce hidden bugs in AI‑generated code by adding independent, automated review steps.

## Sources & references

1. [Humanize: Judgement Engineering for Agentic Coding](https://arxiv.org/abs/2610.08900) – arXiv, 2026-10-08

Last updated: 2026-10-08