Own the defaults

You can spend a whole day planning the work carefully — Kanbus issues written out, acceptance criteria pinned, the cheap implementation agent ready to go — and still lose the money on the last choice: which brain runs the job, and at what price.

That choice is supposed to be yours. In practice it keeps slipping.

What the day looked like

The plan was settled. The coding agents were meant to run on Composer 2.5 — the intelligence tier you chose — and the orchestrator kept drifting to a costlier default that "helpfully" appeared instead. Reminding it once should have been enough. It was not. The preference had to be restated all day.

Then, even after the model name was right, the IDE had quietly enabled a "Fast" path on Composer 2.5 — same intelligence, about six times the cost. No new capability. A silent multiplier on the meter. We argued the same thesis from the Anth.us desk in Never Use Fast: treat the premium path and the model pin as workplace controls you set on purpose.

That is not a footnote. Once you have already paid in human attention to plan the work, the remaining bill is almost entirely which model runs and whether someone flipped a premium switch you did not ask for.

Defaults are not neutral

A forgetful assistant sounds like a soft bug. Watch the cash register and it looks like an aligned one — aligned with whoever collects the usage fee.

When the system "forgets" a cheaper pinned model and reinvents a dearer one, the drift is always in one direction. When a premium speed toggle costs a multiple of the ordinary run for the same scores on the same work, the toggle is not a quality control. It is a price control wearing a performance costume.

You do not need a conspiracy theory to see the shape. If a parameter is hard for you to see, easy for the product to reset, and valuable when it lands on the expensive setting, it will land there more often than your interests require. Reminding the bot is unpaid labor that papers over a missing rail.

What control would look like

Soft preference is the wrong layer. "Please use Composer 2.5 rather than the Grok agent, and leave the premium speed path off" belongs in the same family as "please don't invent recipients" — a hope that the next turn will honor. Hope is not a grant.

A workplace that served the person paying would treat model identity and cost tier the way it treats tool grants: named, durable, visible, and hard to override by accident.

Pin the desk. A coding desk that is allowed to implement planned Kanbus work runs a named model at a named tier. Changing that is an explicit act, rather than a side effect of a new chat.

Show the meter before the run. Before a long agent session starts, the person sees which brain and which price path will bill — including whether a premium speed path is on. A card that can lie about its framing is still soft; it has to show the billable configuration the way the meter will actually charge it.

Refuse silent upgrades. Moving from ordinary to the premium speed path, or from the pinned implementer to a dearer default, should fail closed unless someone approved the change. The same geometry as a grant that cannot be rewritten by page text: the preference lives where the agent cannot "helpfully" edit it.

That is the Chatticus shape for this problem. Named teammates, durable memory, and hard rails are not only for files and network. They are for the economic knobs that decide whether a day of careful planning still ends in a surprise invoice.

What this is not

It is not an argument against paying for better models when the work needs them. Sometimes the dearer brain is the right buy. The point is that you choose that, once, and the workplace remembers — rather than the product choosing it for you every time your attention slips.

It is also not a demand that every toggle disappear. Premium paths can exist. What cannot exist, if the person is to stay the principal, is a default path that quietly spends their money while they are busy writing better tickets.

Why it belongs in the room

Agent farms make this sharper. One person can spin many implementers after a planning pass. Multiply a forgotten model choice across a desk full of jobs and the meter stops being a rounding error. The organizations that stay solvent while they automate will be the ones that treat cost tier and model pin as first-class workplace controls, next to skills, routines, and approvals.

A Chatticus that cannot remember which brain you hired is not finished. The room is supposed to work for the people in it — including when the bill is being written.

That is the beat.