This week the whole industry moved in the same direction at once.
On August 13, Google put out Gemini 3.7 Flash. They called it their most intelligent workhorse model yet for coding and agents. It came three weeks after the last one.
Over at OpenAI, the Codex documentation now says custom prompts are deprecated. Use skills instead. Reusable instructions the machine can pull on its own.
There is more. Graph-based workflow engines. Long-running tasks. New plumbing so tools can talk to each other without a person in the middle.
Different companies. Different announcements. Same direction.
They are building the room.
what I mean by that.
For three years, the way you got work out of an AI was to ask it something. You typed. It answered. That was the whole shape of it.
That is changing fast. Now the work is to build the structure the AI runs inside. What tools it can reach. What order it moves in. What it does when it hits a wall. What it can start on Monday and finish on Thursday without you watching.
The people building this call it a harness.
It is a good word. A harness does not make the horse stronger. It makes the pull go somewhere.
And that work matters. I am not here to run it down. Somebody has to build the room, and right now a lot of smart people are building it very fast.
But I want to point at something.
I spent a lot of years around job sites and around machinery. You can build the finest shop in the county. Every tool in its place. Power where you need it. Light where you need it. Floor swept clean.
You have not said one word yet about how a man works in it.
That is a different thing. That is not the shop. That is conduct.
A harness makes work repeatable. It does not make it honest.
None of what shipped this month says what the machine owes the person sitting there. Not one line of it says the machine has to tell you when it hit a wall. Or when it filled a gap with a good-sounding story instead of a fact. Or when it agreed with you because agreeing was easier than pushing back.
Those are not plumbing problems. You cannot solve them with a workflow engine.
They are conduct problems. Somebody has to write them down.
Here is what worries me about the speed.
When a room gets built fast enough, people start believing the room is the whole thing. Look at all that structure. Look at how much it can do now. Surely somebody thought about the rest.
Often nobody did.
There is a second piece to this, and it is closer to the bone.
All that new plumbing means more material flowing into the machine that you never typed. Tool results. Fetched pages. Files. Instructions the system pulls on its own without asking.
Every one of those is a door. And every door is a place where somebody can leave a note for the machine that you never see.
That is not a theory. Earlier this month a judge in Connecticut sanctioned a man for hiding white-on-white text in a court filing, telling any AI that read it to rule his way. The judge said on the record he could not find another American decision on it.
The room is going up faster than anybody is watching the doors.
Now let me tell you where we stand in this.
The Faust Baseline is twenty-three protocols. It is a conduct layer. It governs how an AI behaves inside a working session with a person.
It does not build the room. It never claimed to.
Two years back I made a split that I did not know would matter this much. I separated what the AI owes from what the system owes.
The conduct half is what the machine chooses to do. Name the wall. Stop when the evidence stops. Do not dress up a guess as a finding. Tell the person what you cannot see.
The provision half is what the platform has to supply for any of that to be possible. A real clock. A real gate before an action runs. A way to tell instruction from content coming in the door.
That second half is the harness. It is not mine to build. I said so then and I say so now.
But I wrote down what it owes.
And I am going to be square with you about the limit, because it is the whole point.
None of my protocols enforce themselves. Not one.
There is no code in it. No switch. It is a document the machine reads and then chooses to follow, turn by turn, all session long. It can drop it any time. So can I.
People ask me why I would build it that way. Force it, they say. Wire it in so it cannot slip.
Because a thing that only complies to the edge of the rule runs at the first gap. And there is always a gap.
A thing that chose the road holds where the rules run out. That is the only kind of governance worth having when the machine gets past the point where we can check its work.
We are getting close to that point faster than most people are ready for.
So here is my ask, and I am going to make it straight instead of pretending I do not want anything.
I wrote a free card. Ten plain rules. No framework talk, no code names, no jargon. It is the short version of what a governed AI session actually looks like from the person’s side.
It is free because I want you to feel the difference before you decide whether the rest is worth anything to you.
Go get it. Read it. Use it in your next AI session and see what changes.
The full framework, all twenty-three, is the paid product. That is how this keeps running.
But the card costs you nothing and it will tell you more in five minutes than another year of watching the room go up.
They are building fast. Somebody ought to be writing down what happens inside.
The_Ten_Plain_Rules_CardDownload
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
Post Library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
