The week after Labor Day, a product team comes back to four agent-generated documents: a competitive scan, a user-research digest, a technical feasibility memo, a draft positioning statement. Each is adequate on its own terms. Together they don't yet say anything.
The work that follows is deciding what they should become. Which findings belong in the executive brief and which get filed. Whether to lead with technical constraints or let the market opportunity set the frame. How much competitive detail the VP actually needs before the quarterly review, given the priorities she signaled in a Slack thread two weeks earlier.
This is composition — assembling agent-produced material into an artifact shaped for a specific reader and a specific decision. It is already widespread. A study of 885 product managers at a large software company found the most common AI workflows involved turning fragments — brainstorming notes, planning documents, peer feedback — into structured deliverables whose shape depended on who would read them and what the document needed to accomplish. Diary research across knowledge workers shows the same pattern from the worker's side: people exported material from several applications into a model, used it to organize disparate formats, then went back in by hand to adjust tone, narrative order, and emphasis for different audiences. The model organized; the person decided what the organized material meant.
I've written before about review-shaped work — the validation layer that grows around delegated execution, where the question is whether the output is correct. Composition sits earlier in the sequence. Before you check accuracy, you decide whether this is the right material at all, arranged in the right order, for the person who has to act on it. Checking accuracy draws on domain expertise. Deciding what belongs draws on something else: the politics, the audience's habits, the strategic priorities that determine which facts matter this quarter and which are noise.
There's evidence the assembly half already occupies more of the collaboration than the generating half. An analysis of 158,000 enterprise LLM conversations at a professional-services firm found editing or improving supplied material appeared in more conversations than generating new content did — 46 percent against 35 percent.
As execution gets cheaper, that ratio should widen. The supply of competent raw material grows faster than the supply of people who understand what any particular stakeholder needs from it. A finance team can produce twenty scenario summaries in an afternoon. The binding constraint is the person who knows which three the board will take seriously, and how to frame the trade-offs among them.
The instinct is to invest in better prompting, on the theory that better inputs produce a better result. But the judgment that turns adequate parts into a usable artifact runs on knowledge the agent doesn't have and can't be given cheaply: what this organization cares about right now, what this audience will do with what you hand them, what should be left out.
What seems underappreciated is how invisible this skill was before agents separated it from production. When you wrote the analysis yourself, that knowledge was bundled into the act of producing it. You chose what to emphasize while you were still working out what to say; the selecting and the making were one motion. Cheap execution pulls them apart. The selecting turns out to be the harder half, and it doesn't improve when the models get faster. Most organizations have no category for it — no title, no measure, no development path — which means they are unlikely to notice when it becomes the thing they're short of.