Private-agent productivity.

Leaders keep asking me how their company should start adopting agents. The question is about two years late. In every company I talk to, adoption has already happened: privately, one person at a time, and mostly without anyone deciding it. The real question is not how to start. It is what to do about the adoption you cannot see.

The result is private-agent productivity: real individual gains from agents that never become organizational capability.

A note on evidence: this is a generalization from conversations with teams and from years spent building collaboration systems, not from a survey, and I have deliberately avoided inventing numbers to make it sound more rigorous than it is. It also has a clean falsifier: if your company's agent use is visible, shared, and discoverable, this essay's argument does not apply to you. I have not yet met that company, but I would genuinely like to.

The mechanism starts with a fact worth respecting: private adoption is rational. The tools are single-player. Sharing has friction and no reward; nobody gets promoted for demonstrating a prompt in a meeting. So a developer runs a coding agent all day, a product manager tunes a research assistant, an operations lead builds an automation, an executive reads a private morning briefing, and each of them is right to do it. We have also seen this movie twice before. Spreadsheets entered companies through individual desks, and SaaS entered through individual credit cards, before either became organizational infrastructure. Capability that arrives through individual choice always precedes the structures that make it compound.

The costs are structural, and they grow with adoption rather than shrinking. The same work gets done repeatedly because nobody can see it was already done. Context gets reconstructed from scratch in every session. The best agents in the building are undiscoverable outside the person who made them. Output arrives without sources, so trusting it means redoing it. And an organization cannot govern usage it cannot see, which turns every serious question about AI risk into an archaeology project.

A product manager, a founder, and a sales lead each run a competitor analysis through their own agent in the same week. Three sets of sources, three slightly different conclusions, zero awareness of each other. The strongest analysis lives in a personal chat history and never surfaces. The decision gets made on the weakest one, because it was the one pasted into the meeting doc. No individual made a mistake anywhere in that story. The organization still reasoned worse than any one of its members.

The governance-first objection deserves its strongest form: some data should never enter external tools, policy exists for good reasons, and "let a thousand agents bloom" is not a security posture. All true. But a ban does not remove usage; it relocates it further from view, and visibility is the one resource governance cannot function without. The second objection says vendors will add sharing features and the problem resolves itself. Sharing a transcript is an artifact, not a capability. What compounds is identity, attribution, reusable context, and discoverability, and those are properties of a shared environment, not of an export button.

What is happening now is broad, private, growing agent adoption. What I expect, with more confidence in the shape than the schedule, is that the gap between individual and organizational returns becomes the defining complaint of enterprise AI: individuals love the tools, and the organization cannot point at the value. What to do about it is unglamorous: make agent work visible before trying to optimize it. You cannot reuse, improve, or govern what you cannot see, and every quarter of invisible adoption makes the eventual consolidation harder.

Organizations exist to compound individual insight across people and time. Private-agent productivity breaks exactly that step.

The people adopting agents privately are not the problem; they are the early signal of where work is going, and they found leverage the organization had not offered them. The fix is the oldest tool teams have for work that matters: give it a shared place, where contribution is visible, credit is attributable, and the next person can build on what the last one learned.