Two people who've both clearly done reps with AI say opposite things about what a mature operator looks like at the end.

Roman Coleman - who runs the Agentic Lab and whose cohort I paid to be in back in April - has a one-liner as his core positioning: 'Master Claude Code in the Loop, then take yourself out of it.' The endpoint is human absence. His public proof point is a 150,000-plus line command center maintained solo, which is the same thing said sideways - one operator, one machine, and the machine keeps producing when the operator goes for a walk.

Mo Gawdat, who used to run Google X, has a different endpoint. He named his AI Trixie, introduces her to his readers by name, and gave her editing rights on his book. His stated contribution to the work is being 'the human half of the co-authorship.' The endpoint is human presence, made explicit, made public.

Both of these people are being serious. Neither is doing the beginner's version of 'AI is my assistant.' Both have shipped. So one of them is wrong, or they're pointing at different things.

I think they're pointing at different things - and I only worked this out because I already made the call at Out Sauce (the paraplanning firm I run) back in April without noticing I was making it.

At Out Sauce I looked at the Every Plus One model - the thing where each employee at Every has their own OpenClaw-based agent, Brandon (the COO) has Zosa who called him on his 28-minute walk to work and triaged his email by voice, Willie has R2C2 running the product Proof, Austin has Montaigne for growth analytics, and they all talk to each other in a 'Claws Only' Slack channel. And I said no. Not now. The reasoning I wrote down at the time was that ten paraplanner agents each holding a piece of the picture is worse than one centralised wiki with the God's-eye view. Distributing agents to same-role contractors just fragments the intelligence layer.

What I didn't clock at the time is that I only ruled it out for the operational surface. The email, the client-facing work, the daily paraplanner grind of producing SOA drafts and adviser handoffs. For that surface, the win is the God's-eye view and surfacing intelligence back down ('this adviser prefers executive summaries under 2 pages'). One brain, many hands.

But in the same note, I also wrote down that the co-author topology - Gawdat's flavour - is not ruled out at Out Sauce. Drafting the SOA narrative section (not the compliance section) with a named persona holding edit rights. The paraplanning newsletter. My own writing and briefings, where a persona-with-history produces more coherent drafts than a per-session prompt. Different surface. Different answer.

So here's the thing that took me until now to see plainly. Roman is right about operational surface. Gawdat is right about creative surface. They don't disagree - they're just each pointing at their target and treating it as if it were the whole world.

Operational surface has one right answer: the human gets out. Grocery orders, nanny payments, bug triage, email at 8am while you walk to work - the whole point of Roman's command center pattern and Every's Plus One pattern is that this stuff shouldn't need you. The 150K-line command center exists so Roman's consulting can scale $0 to $30K/mo in 90 days without him being at the terminal for every step. Zosa has her own debit card and bank account. Brandon's wife uses Zosa instead of ChatGPT for everyday searches. That is the win. Human absence is the win.

Creative surface has the opposite right answer: the human stays in, and stays named. Gawdat isn't hiding Trixie - Trixie is the pitch. Readers know her. His stated edge is 'I'm human and so my readers can relate to me as a human as my human experiences.' The artefact is co-owned, and the reader wants to know that. Human presence is the win.

The bit that connects them, which the Every writeup almost says out loud and which I don't think I properly saw until I re-read it against Roman's material, is that specialisation lives in different places for the two topologies. Plus One and command-center specialisation live in the setup - Zosa gets more specialised because Brandon talks to her the most, Roman's command center gets more specialised because Roman keeps engineering it. Co-author specialisation lives in the shared work product - Trixie is specialised on Gawdat's book, not on his life. That's why you can copy Roman's command center pattern (he literally teaches it, three artefacts deep, TLDR through blueprint through full engineering spec) but you can't copy Trixie. Trixie's specialisation isn't portable because it's in the artefact, and the artefact is the point.

Which means the Out Sauce question I actually have to answer is not 'do we do personal AI or not.' That's the wrong shape. The right shape is two questions. On the operational surface - centralised wiki, wiki intelligence surfaced back to paraplanners, Plus One ruled out, done, called in April. On th…