Human-in-the-Loop Is You: Orchestrating Multiple Coding Agents

“Human-in-the-loop” is supposed to mean a person makes the judgment calls a model can’t. In practice, with Claude Code, Codex, and Cursor all open at once, it usually means something dumber: copy-pasting the same bug report into three chat windows, re-explaining context each time, and losing track of which agent already tried which fix.

This workshop puts the alternative in your hands. We’ll work in one small, real codebase and drive several coding agents against it through Omnigent (https://omnigent.ai/), an open source meta-harness built that sits as a common layer over Claude Code, Codex, and agents you write yourself.

You’ll go from three separate agents to one workflow. Along the way you’ll orchestrate two agents on the same task and have them debate an approach, set a stateful spend cap and watch it halt a tool call mid-run, write a contextual policy that blocks a risky command instead of babysitting every action, and share a live session by URL so the person next to you can review and co-drive rather than watching you screen-share a terminal.

You’ll leave having actually run a multi-agent workflow end to end, with a set of patterns you can take back to your own repositories. Expect a few things to break; that’s part of the lesson on where this approach works today and where it doesn’t.

You’ll need a laptop, Claude Code and Codex installed and a GitHub account to participate.