A lot of people ask me how I'm shipping products so fast lately. The honest answer is that I almost never write code from scratch anymore - I orchestrate two coding agents that are each really good at one specific thing. Here's the workflow that's been working for me.
The split: one agent thinks, the other builds
The mental model is simple: Codex plans, Claude Code executes. I treat them like two coworkers with very different strengths, and I stopped trying to make either one do the other's job.
- Codex (GPT-5.5) is my planner. I use it to think through architecture, break down features into steps, sanity-check my approach, and catch the "wait have you thought about X" stuff before I write a single line of code.
- Claude Code (Opus 4.7) is my builder. Once I have a plan I trust, I hand it off and let Claude actually do the implementation - edits, refactors, test runs, the whole loop.
GPT-5.5 in Codex is insanely good at planning. It thinks slowly, considers edge cases, and pushes back when something doesn't add up. Opus 4.7 in Claude Code is insanely good at executing - it stays disciplined, follows the plan, and gets the file edits right the first time.
What this actually looks like in practice
When I'm building a new feature for a product, my flow is usually:
- Yap at Codex. I just dump everything I'm thinking - the feature, the constraints, the half-formed ideas, the stuff I'm worried about. Codex turns that mess into a real plan with concrete steps.
- Push back on the plan. I read it carefully and challenge anything that feels off. Codex is good at incorporating pushback and revising. Two or three rounds of this and the plan is usually solid.
- Hand the plan to Claude Code. I paste the plan in, point Claude at the right files, and tell it to implement step by step.
- Let Claude cook. It edits, runs tests, fixes issues, and reports back. I review the diff, push back where needed, and ship.
The win here isn't speed for speed's sake - it's that I'm always thinking at the right level. When I'm with Codex, I'm thinking about what and why. When I'm with Claude, I'm thinking about does this diff look right. I never have to context-switch between architect-brain and reviewer-brain mid-task.
Why two agents instead of one
I tried doing everything in one agent for months. Both Codex and Claude can technically plan and execute. But there's a real cost to making one tool wear both hats:
- The planner-mode of a builder agent is rushed. It wants to start writing code.
- The builder-mode of a planner agent is too theoretical. It explains instead of edits.
Using each agent for what it's actually best at turned out to be way more effective than picking a favorite and forcing it to do everything.
What I'd tell someone starting this workflow
- Don't skip the planning step. It's tempting to jump straight to Claude Code because it's so satisfying to watch it ship. But 20 minutes of yapping with Codex saves hours of rework.
- Treat your agents like teammates, not magic. Push back, ask "are you sure", challenge the plan. They get better when you treat them seriously.
- Pick one builder and stick with it. Switching builders mid-feature is chaos. Claude Code Opus 4.7 has been my daily driver and I haven't found a reason to swap.
This workflow is still evolving - I tweak it almost every week. But the core split (Codex plans, Claude builds) has been the highest-leverage change I've made to how I work this year. If you're shipping products with AI agents and not getting the speed you expected, try splitting the brain in two.
More to come on the specific prompts and project setups I use. Reach out if you want to compare notes.