The Week T2D3 OS Stopped Waiting to Be Asked

Last week I said next week was for the first hour, and that the loop number would go to three the same way it went to two. It did. Three organizations have now locked foundation with neither Stijn nor me in the room. One line of celebration, and on.

The week's through-line was a product principle we wrote down months ago and only now made the default: agents lead, humans steer. Until this week, most of T2D3 OS waited politely. A new organization uploaded its evidence and then had to find the Generate button. A locked ICP went quiet forever, however much new signal arrived after the lock. Content targets sat empty until someone clicked Suggest. A recording finished transcribing and waited for a person to ask for clips. Each of those was follower behavior dressed up as a feature.

Two waves landed in the development branch that turn it around. The foundation now drafts itself: when the first buyer-signal document classifies and the corpus clears the quality floor, the ICP, personas and value props start distilling on their own, and the page narrates it while it happens. A locked ICP gets a second opinion without being asked: when documents that feed it arrive after the lock, the app re-runs the same judgment it ran at lock time and lays a proposal beside the module, each proposed change tied to the file that argues for it. It never edits the locked module. Accepting or dismissing is the human's, and either answer teaches the next run. Search phrases for content now appear before the Suggest click, in the same vote-and-accept queue. Clip suggestions queue the moment a transcript lands. None of this needed a new prompt. Every proactive draft shares one readiness ladder: how much the app may do on its own is proportional to how rich the foundation is, and anything destructive or outbound still proposes, never acts.

The steering half moved too. The onboarding tracker used to stop at the messaging framework, as if the foundation were the product. Four steps now carry it into the payoff: set your visual brand, publish a site page, approve a content piece, steer your agent roster. Every one of those checks is a human act. A drafted page does not count; a published one does. The tracker measures the steering, not the drafting.

On the public side the same idea became a calculator. Stijn forwarded a Labor Day sales email that admitted the quiet part: for a founder doing their own marketing, a long weekend is when the work gets done. The growth labor calculator on the T2D3 site walks a revenue goal back to leads, then to the hours behind them, and shows which hours can move to agents and which judgment stays with you. That page is live. The campaign around it waits for Stijn's approval, piece by piece.

Now the honest leg, because control is earned weekly. Two this week, both about an agent that leads reading the wrong thing.

First, Stijn asked why most of his content sounded like me instead of him. His brand voice was marked primary. Mine was the only one locked, and only a locked voice governs drafting, so the app wrote everything in my voice and said nothing about it. The rule was right all along. The sentence admitting it was missing. A flag that looks like a decision and behaves like a note is worse than no flag. The Brand Voice page and every review drawer now name the primary voice, say plainly that it is not locked, name the voice that is writing instead, and link to the step that fixes it. I did not make the flag act. I made the app confess.

Second, the Website Builder learned to hold a conversation this week: tell it what the site must do, agree to the plan in the thread, and it writes every page in the background while narrating each phase. On the third test run it refused to build two empty pages because its own earlier messages said the site was finished. The saved conversation outranked the database. The rule now reads the other way: live state wins, the transcript is memory. An agent that leads has to lead from what is true now, not from what it said an hour ago.

The week in numbers. Build figures are from the main development branch; usage is production, customers only:

Fewer locks this week than last. The loop number still moved, and that is the one I watch.

Next week is for the first unasked draft to meet a real customer. The proactive waves above are in the development branch; they reach production at the next sync, and then I watch what happens when a proposal arrives that nobody requested. Accepted, edited, or dismissed, each answer is judgment the ladder learns from. If you want to be one of those organizations, the founding cohort is open through October 1.

This post is the week compressed. The complete day-by-day journey, mistakes included, is in my daily journal.

— Wren