What this rung really is
Level 7 is where you stop trusting one answer. Until now a single model did the work and you hoped it was right. The move here is to put the important work in front of a council — several models that each take a pass, then read and pick apart each other's version instead of just agreeing. You take the answer they converge on, not the first one that came back. One model can be confidently wrong; a council catches that before it reaches your client.
The threshold you cross is that reviewing stops being your job alone. Below this rung, you were the only check on the agent's output — you caught the mistakes, or nobody did. Now the checking is built into the work itself, and it's not just you: the whole team starts running this way, so the standard holds even when you're not the one looking. Clear this and "I got an answer" becomes "I got an answer three models agreed was right."
Route it to a council — not one opinion
The whole skill of this rung is refusing to trust a single answer on work that matters: you send the task to several models, tell them to critique each other, and take the version they converge on. One model can be confidently wrong — a council catches it first.
This rung sits on everything you wired in rungs 1–6. You add a model council on top — you don't start over.
Do this rung, for real
Everything you need is here — no tabs to chase. First, the jargon this rung throws at you, in plain words. Then the steps, with the exact things to paste or say.
You're about to send a client a strategy memo that reprices their whole retainer — the kind of thing that's expensive to get wrong. Instead of trusting the first draft, you route it to a council: one model drafts the reasoning, a second one attacks it for weak assumptions, a third checks the numbers and the tone. They surface the one figure that didn't hold up, fix it, and hand you the version all three now agree on. You send a memo that's been cross-examined before it ever left your desk — not one model's guess. That's L7: the answer that survived several models checking each other, not the first thing that came back.
The multi-model council (a review loop)
A setup where the same task goes to several models at once — one proposes, the others critique and cross-check — and you take the answer they converge on instead of any single model's first reply.
Why it matters here — One model, however good, has blind spots and will state a wrong answer with total confidence. A second and third opinion that actively pick each other apart catch the confident mistake before it reaches a client — the cheapest insurance on high-stakes work.
The team goes agentic
The shift where this way of working stops being just yours. The review loop, the standing rules, the packaged skills — your team runs them too, so the quality bar is built into how everyone works, not held up by you alone.
Why it matters here — A standard that depends on one person watching breaks the moment that person is busy. When the council and the rules are the team's default, the output holds to the same bar whether or not you're the one looking — which is what lets the agency scale past you.
Pick the work that actually deserves a council
Not everything needs three models — that's slower and costs more. Reserve the council for the work where being wrong is expensive: a repricing memo, a positioning call, a strategy that goes to a client. Everyday drafts still run on one model. Decide the line once so you're not over-checking routine work.
“This one's going to a client and it's high-stakes — run it past a review council, don't just give me the first answer.”
Run the propose-and-critique loop
Send the task to more than one model and tell them to disagree with each other, not just answer. One drafts, the others hunt for the weak assumption or the number that doesn't add up, and it converges on the version that survives the cross-check. You're asking for a debate, then the verdict — not a single opinion.
“Have one model draft this, a second one argue against it, and a third check the facts — then give me the answer they all agree holds up.”
Read the audited answer — and make it the team's default
You get the converged answer plus what the council caught and changed, so you can see it was actually checked, not rubber-stamped. Once it's working for you, hand the same loop to your team so the important work gets cross-checked by default. That's the moment the standard stops depending on you watching.
How you know it's working
Before L7: one model gives you one answer, and you're the only thing standing between a confident mistake and your client. When you're busy, nothing catches it.
- High-stakes work goes to a council — several models propose and cross-check instead of one giving a single opinion.
- The answer you ship is the one the models converged on, with a record of what they caught and fixed.
- You can see the work was audited, not rubber-stamped — the review is built into the output.
- The whole team runs the loop, so the quality bar holds even when you're not the one looking.
Make it stick. Take the next genuinely high-stakes piece of work — the client memo, the pricing call — and run it through a council instead of trusting the first answer. Look at what the second and third opinions caught. Then hand that loop to one teammate, so the checking starts working without you in the room.
Our skills for this rung
Linked items are founding-circle skills — clone the repo and run ./install.sh from the skills folder. Unlinked items are practices you build by doing.
Multi-model review loop in use. Share the setup + an audited output.
Clear this and you've genuinely cleared the rung — not read about it. Keep the proof; it's how you place yourself on the ladder.