Should we plan with Codex, then code with Claude? Or should we plan with Claude, then code with Codex?

How about yes.

I found myself converging to this workflow for more complicated features. I would give the same prompt to Codex and Claude and ask each to create a PRD file for the feature.

The independence mattered, because the moment one model sees the other's plan, the answer collapses toward whoever spoke first.

Once I had the PRDs, I would manually ask both Codex and Claude to critique the other plan, and revise theirs based on the findings of the other. Eventually I would converge to a final PRD where both models had reached agreement.

This final PRD was meaningfully better than what either model produced alone.

Doing this manually was annoying, so I packaged it.

Presenting The Order of the Agents.

The Order of the Agents convenes a sworn fellowship of AI agents (Codex, Claude, and other CLIs you trust) around a single question. Each agent takes a position, challenges the others, and revises in turn, until the Order issues a final decree. Every oath, critique, and revision is recorded as Markdown, so the reasoning behind the decision is auditable, shareable, and yours to keep.

You can even do a grill-me style intake first to finesse the requirement before the Order convenes. That mode was inspired by Matt Pocock's grill-me skill.

You can install it locally with npm:

npm install agent-order

Requires codex and claude CLIs (or other agent CLIs you trust) already installed and logged in. Use it like this:

npx agent-order@latest "Research and draft a PRD for adding SSO to our app"