Saying what you want, and having it happen
A model can tell you how to update a record. The harder, more useful thing is for it to update the record: correctly, with your permission, and with a trail of what it did and why.
When you ask for something, the platform turns your request into a structured specification of intent - not a paragraph of prose, but a formal statement of what you want, what it touches, how it should be verified, and the reasoning behind it. That specification is the source of truth; it is designed so that the concrete work can be generated from it deterministically, the same way twice, rather than improvised.
From there it becomes a real action. The platform will create, update, delete, or submit an actual record in the live system, a task or an invoice or a booking, not a suggestion you then carry out by hand. Two things guard that. Nothing runs on a single turn: the platform shows you exactly what it is about to do and waits for you to confirm it on a separate step, so an action is never a side effect of a sentence. And every write is bounded by permission; it can only do what the person asking is allowed to do, never more.
What you are left with is a record on both sides: the change itself, and the specification that asked for it, kept together. If you want to know why something happened, the intent that caused it is still there to read.
This is the piece that turns a capable talker into something that can be trusted to act. Most of the discipline in the platform lives here, at the boundary between a request and a change to the world.
The same discipline points somewhere we did not expect. Decompose a hard problem into typed steps and hand the deterministic parts to a solver rather than the model, and a small model that runs on local hardware reasons far more reliably than one left to think in a single unstructured pass. We measured this on a hard reasoning benchmark: the structured, solver-assisted pipeline beat flat model "thinking" several times over at the same model size, and held up better when the problem was perturbed. Advanced reasoning turns out to be less about a bigger model than about the structure you put around it.