The council reviews it before I do
Most plans and changes my agents make get reviewed by three models in a loop before I look at them. This is why I use three, and the rules that stop the loop from growing the change.
Over this year, as my velocity increased and more and more code was generated each day, I started struggling to keep up with it and do adequate review, so I started throwing more compute at the problem. Now, for most plans and changes my agents make, I don't want to be the first one to look at them. I want multiple models to review the work first, and I want the agent to work off those reviews in a loop until it's in a good place. Then I'll spend time looking at it, and if necessary send it to a teammate for a second review.
It's a common thing used by many devs now, it's called a council. In my setup it's a global skill that every agent on every machine follows. The agent brings in three sub-agents, each a strong model at high effort, and has them review what it did before it reports back to me with a rating out of 10. It started out as a single model review loop, but after some experimentation I got MUCH better results with a set of models.
49 councils. 36 took 1 round, 9 took 2 rounds and 3 took 3 rounds. Then the 3-round limit. 1 council took 6 rounds, in October, before the limit.
The council goes first
I went back through the T3 Code records on all of my machines. Since late October I've run 49 councils, with 69 rounds of review between them. 36 of the 49 were done after one round, and a round takes about five minutes.
So most of the time the council adds at least one round of review before a change gets to me. By then three reviewers have read it and the agent has fixed what it agreed with. What I get is the diff, a score from each reviewer and a short list of what the agent declined and why.
Why three models
In my experience a council beats a single reviewer because different models and model families are good at different things. Leaning on a few of them gives a lot more variety, and more eyes on the changes in general, so it's more likely to catch issues.
My council right now is Claude Fable 5.2, Claude Opus 5.5 and GPT 6.1 Sol. GPT 6.0 Astra had a seat until recently, and I've tried Muse as well.
The records show the variety. In the rounds I looked at, 19 turned up a serious finding, and in 8 of those only one of the three reviewers raised it. It wasn't the same model each time. Sol was the only one three times, Fable three times and Astra twice.
Here is one of them. On a Mac app, both GPT reviewers approved a change at 9/10. Fable gave it an 8 and asked for changes, because an error message said the app was falling back to the default microphone when there was no microphone at all, and it logged that on every retry. The agent agreed and fixed it.
Interestingly, they grade differently too. Often when one reviewer alone gave the lowest score in a round, it was a Claude model about 80% of the time.
When the loop doesn't stop
There is a catch with a review loop though. It is that reviewers will generally always find something. Often it's not big but things often will show up.
In October, before I added a round limit to the skill, I asked an agent to write a plan and run a review loop until it was happy with it. The draft was about 490 lines. Six rounds and almost an hour and a half later it was over 1,000 lines, and about a fifth of that was a log of the reviews. By rounds 4 and 5 the reviewers were mostly finding problems in the fixes from the round before. Round 6 was meant to be blockers only and it still found two, and those were fixed and never reviewed again.
Left alone, the same loop can balloon a hundred-line change into a 3,000-line one.
The rules that stop it
A lot of the skill is now about saying no. Here is a trimmed part of the loop version.
Reviewers advise. You decide. Declining a finding is a normal outcome, not a
failure [...] The default answer to anything outside the intent is no.
A small, correct change that ends at 8.7 beats a bloated one at 9.3.
- Default: 3 rounds total. [...]
- Extend to 5 only for security, permissions, auth, data loss, or
state/concurrency [...]
blocker = intended case wrong. should = real bug callers will hit, in intent.
nit = rest.
Three members agreeing on a nit still makes it a nit, and you can decline it.
If review grew the diff, cut first. Do not agree just to end the loop.
Every finding gets a label. The agent declines a nit in one line and lists it as a follow-up, so I can still see it, and a declined nit can't start another round. The target is a 9/10 average, but a smaller change with a lower score wins.
Since the limit went in, no council has gone past three rounds without explicitly asking me to continue.
This is the score table from one of my pull requests.
| Reviewer | Round 1 | Round 2 | Round 3 |
|---|---|---|---|
| Claude Opus 5.5 | 8/10 | 6/10 | 9/10 |
| GPT 6.1 Sol | 10/10 | 8/10 | 10/10 |
| GPT 6.0 Astra | 10/10 | 8/10 | 10/10 |
Round 2 went down because all three caught a real regression. During the round 1 fixes, a test rewrote the lockfile without one of the dependencies and the agent committed it, so a frozen install failed. The agent restored it and round 3 approved.
That's the order now. The council reviews a change first, and I review what comes out of it. No more time wasted reviewing code straight from the model.