Trying a Setup Where Claude Fable and ChatGPT Sol Review Each Other

Coding agents are a lot of fun. Lately I’ve been burning absurdly capable minds — Fable, and Sol over at Codex — on my own small projects, and enjoying the pace of the progress.

The rest of the world is well ahead, though. People have agents working through the night, and everyone is talking about harnesses, loops, graphs. I like to follow a trend. So as a start, I had Claude and Codex build me a setup where they review each other’s work, run in loops, and write the things I want to read.

Here’s what we ended up with.

• Research runs on two independent tracks. Fable and Sol investigate the same question without seeing each other’s results, then the findings are merged. If either one flags a claim as important, it can’t be quietly dropped in the merge (this is enforced by a check, not by a prompt).
• The draft is polished by a loop between the two. Sol reviews it (blocking / quality / minor), Fable applies the fixes, and it goes around until there are zero blocking issues, with a cap on the number of rounds.
• Three human gates: approve the direction, mark up the draft, decide whether to publish. Publishing is never automated.
• This run: 9 review loops, 32 blocking issues resolved, 4 rounds of my own edits. Twice, one of my comments led an agent to go do more research and find an error in its own earlier text.

What I learned is that having them review each other still isn’t enough.

Research and sourcing get done convincingly well. That’s the good part. But the structure and the presentation stay blurry and long-winded. And the prose keeps its unmistakably AI-ish habits — pointless metaphors, that em dash — no matter how firmly I tell it to stop. Unlike coding, where people broadly agree on what good looks like, “good writing” varies from person to person. Maybe that’s why it’s hard.

The odd part is that the tone it uses when explaining things to me in a normal chat would be perfectly fine. Something degrades the moment it starts writing an article meant for publication.

I was going to put the resulting piece here as a sort of burial, but publishing something whose accuracy I can’t vouch for felt wrong, so it’s scrapped.