AI Coding Weekly

Theo's council skill makes three models argue before a merge

Cross-model review as a merge gate, with no numbers yet on what it catches or what it costs.

Theo says he wrote a "council" skill that makes his agent argue with Opus 5.5, Muse Wonder Full and GPT-6.1 Sol before it is allowed to merge. That is the whole disclosure so far. He has not said what the council has caught, how often it blocks a merge, or what three frontier models per merge costs him.

Theo - t3.gg
@theo
X
I wrote a "council" skill that makes my agent argue with Opus 5.5, Muse Wonder Full and GPT-6.1 Sol before it is allowed to merge
Oct 23, 2026 · View on X

What is actually claimed

One engineer, one setup, one sentence. The skill is a gate, not an adviser, which is the part worth noticing. Plenty of people already paste a diff into a second model and read the reply. Wiring a disagreement step into the merge path means the agent cannot close the loop on its own work, and it means the gate fires on every merge rather than on the ones a human happened to feel nervous about.

What nobody has yet is evidence. Theo has not posted a single example of a bug the council stopped, and without that there is no way to tell whether three models arguing finds real defects or mostly generates review theatre with a token bill attached. He has previously run models against each other on a single bug and reported which fix he thought would not have worked, as in his three-model test, so the comparison habit is not new. Turning it into an automatic gate is.

Why the pattern keeps coming back

The volume numbers people post make the reason obvious. Theo has claimed 245 PRs in 24 days. At that rate human line-by-line review stops being the thing standing between an agent and main, which is roughly the argument Addy Osmani made when he said line-by-line review is dead as a gate but the gate should stay. Something has to sit in that slot, and other models are the cheapest thing available.

For working engineers, the open question is whether three models trained on overlapping data disagree about the things that actually break production, or only about style. Until someone publishes a catch rate, a council skill is a plausible idea with no receipts behind it.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.