A new open-source plugin called Hubo adds a two-agent implement-and-review loop to AI coding tools, so that every change is written by one agent and independently critiqued by a second before it reaches you. It installs into Claude Code, the Codex CLI, GitHub Copilot CLI, and OpenClaw, and ships under an MIT license.

What This Enables: A Built-In Code Reviewer

Hubo splits work between two persistent roles. A work agent implements the change, tests it, and defends its decisions. A review agent independently inspects the diff and raises numbered findings. The two iterate against each other until every concrete finding is resolved, and only genuine decisions that need human judgment surface to you. Instead of accepting a single agent's first draft, you get output that has already survived an adversarial pass. In Claude Code you invoke it with a slash command such as /hubo implement optimistic updates for the cart.

Why It Matters for Builders

The weak point of autonomous coding agents is that they mark their own homework. A single agent that writes and then approves its own diff will happily ship a plausible bug. Hubo formalizes the reviewer-implementer split that experienced teams use in pull requests, and runs it automatically inside the agent loop. For anyone shipping real code with Codex or Claude Code, that extra review pass is a low-cost way to catch mistakes before they land in your working tree.

Key Details

License: MIT, open-source.

Supported tools: Claude Code, Codex CLI, GitHub Copilot CLI, OpenClaw, and standalone via skills.sh.

Model: A work agent and a review agent spar over each change until findings are resolved.

Status: Posted to Show HN on July 25, 2026 as an early but usable project.

What to Do Next

If you already run one of the supported agents, install Hubo and trigger it on a small, well-scoped task to see how the review agent's findings read. Because it is an early release, keep your own review in the loop rather than trusting the two-agent output blindly, and use it to raise the floor on agent-written diffs rather than to remove yourself from the process.