Your agents shipped the pricing you dropped two weeks ago

TL;DR: DevDay gave agents more to do and nothing told them what for. The missing layer is a place for the decision the work serves, so capable agents keep building, competently, against a version of your company that no longer exists.

OpenAI's DevDay this week was its biggest, by its own count more than twenty announcements. The ones that matter to a small company are the ones about work getting done without you: Dots, always-on agents that take work off your plate; team tasks that run weekly updates on a schedule or when an email lands; a plugin that watches a project board and drafts a plan when a task appears; Codex in the cloud from your phone, reviewing pull requests before you wake up; a Decisions API that picks an agent's next action from a list you supply; ChatGPT Space, one place for your team and your agents to share knowledge; Pages, documents people and agents write together.

It's a good set of tools and we will use several. If you run a company of five, the plans drafted while you sleep will be competent.

Monday morning

You open the Space. The dot drafted the plan for the task that appeared Friday. The weekly update is written. Three pull requests reviewed, one merged. The Pages are longer than Friday, and they're good.

None of it is wrong, exactly. The plan assumes you're still going after the mid-market segment you leaned toward in July. The pricing in the update is the one from the deck. The merged pull request builds the onboarding flow for the per-seat model.

You did write down that you stopped believing in per-seat. Two weeks ago, in a Page, under a heading, in a document that has grown since. The agent that built the onboarding flow read the pricing spec, and the spec still said per seat, because nothing marks a spec stale when a Page changes. The Space holds what everyone wrote. It doesn't hold which of it is the company's current position, what that position rests on, or what it would take to retract it.

So the work was done, on time, by capable agents, against a company that no longer exists, and the sentence that would have stopped it was in the same workspace the whole time.

The pain is not output

Founders of small companies don't run short of output anymore. A year ago the problem was getting things made. Now the drafts arrive before you ask, and the question that keeps you up is: is any of this pointed at the right thing, and how would I know before the quarter is over?

Every tool announced this week answers "what did we say", "what happened" and "what should the agent do next". None answers "what are we betting on, and is it still true". A workspace accumulates knowledge, a page accumulates text, a task runs on a schedule. The decision, the thing a company actually turns on, is in there somewhere as prose, with no way of saying which other pages it changed.

The cost has a date on it. The day you learn the segment was wrong, the finding arrives with six weeks of competent work built on it. The agents didn't fail. Nothing told them which page was the decision, and nothing could tell them the answer had changed.

What the agents read first

StratOS is a record of what your company has decided, and it's what the agents read before they work.

You commit a position: this customer, this price, this reason. The record keeps it with what it rests on and the number that would prove it wrong, by a date. When you supersede it, every read and spec that rested on it is marked stale with the reason, by the record, not by someone remembering to edit the spec.

Your agents, in Codex, in a dot, in Claude Code, read the record before they build; a superseded decision is refused before the first line, not caught in review. When the number comes in, from you or from the agent that shipped the thing, the record scores it against your threshold and asks you for a verdict. It never decides for you.

The outcome on Monday morning: the weekend plan was drafted against the segment you're actually betting on, because that's what it read. The pull request for the abandoned pricing was refused before it was written, because the spec it would have built from was marked stale the day you changed your mind. When the per-seat model fails the number you gave it, the record puts the verdict in front of you on the Tuesday it happens, not at the quarterly review.

We didn't build the agents. OpenAI, Anthropic and a dozen others build better ones every quarter, and this week's are very good. We built the part that says what they're for, and the part that tells you early when you were wrong about it.

Run the dots. Point them at the record first, and the plan waiting for you on Monday will be built on the segment you still believe in.