Revise Blog - Building a post-AI word processor #
I’ve been practicing pair programming with my coding agents, and I don’t see this coming up in the discussions around “AI coding meta”. People compare models, and exchange notes on effort levels, but in general the discourse around AI coding is focused on choosing a single model and letting it work.
When I started my career working in the Codecademy office in NYC, we would commonly sit together at the same desk, appoint a driver and navigator, and work on problems together. This has benefits for humans that AI models won’t enjoy, like socializing and learning, but it also results in better output. I think for mission-critical work, having AI agents pair together is a no-brainer.
How? It’s easy. When you’re giving Codex instructions, tell it:
“Keep a running session with Fable 5.1 via claude with my Claude Max account. Pair program with Fable on this task. You are driving, Fable is navigating. Have Fable review and give feedback on any architecture changes you intend to make, and code changes you implement. If you and Fable don’t agree about something, discuss the problem together until you do. Neither of you is in charge; Fable is your equal peer.”
You could put this into AGENTS.md, use it as a skill, or whatever flavor of prompting you prefer.
In my experience, even when running Astra xhigh, Codex will get multiple review rounds from Fable on something before it moves on, and Fable catches legitimate issues constantly. It tends to look more broadly, helping Codex escape its tunnel vision. The results are generally higher quality and more thorough than either of these models working alone.
Of course, you can reverse these roles, but in my experience Codex’s usage limits are much more generous and so having it drive makes sense to me.
And yes, it costs more to work this way, but you get what you pay for. Two heads are better than one, as they say!