Think about a baseball umpire calling balls and strikes. The technology to do that job automatically has existed for years, and it is more accurate than any person. Almost the entire task is now machine work. And yet a human still stands behind the plate, because the last small fraction of the role, the authority, the ceremony, the willingness of everyone to accept the call, has not been handed over. The job is ninety-nine percent automatable and zero percent automated.
A lot of knowledge work is arriving at that same place with agents, and the umpire is the honest picture of why. The blocker is almost never the technology. It is the last stretch of a job that someone is not ready to give up, and a structure built around the way the work used to be divided.
The Last One Percent Is Human
When I say most of a role can be automated, people hear a threat. That is not how it plays out day to day. What actually happens is that the mechanical majority of the work becomes cheap, and the whole thing hangs on a thin remainder that is judgment, trust, and permission. The agent can do the ninety-nine. It cannot grant itself the one.
This is not a criticism of anyone. It is just the terrain. Holding onto that last piece feels responsible, and sometimes it is. But it is worth being honest about when the holdout is genuine judgment and when it is only a hand that will not let go of the wheel.
The Year That Took a Week
The clearest lesson I have on this came from a piece of work that everyone had agreed was roughly a year away. It was on the roadmap at that distance because, built the traditional way, across teams and backlogs and the gaps in everyone's calendar, that is honestly what it cost. It had been quietly filed under too hard for now.
I built it in a week, with an agent, mostly out of frustration. I want to be careful with that sentence, because it invites a lazy reading, which is that the year estimate was foolish. It was not. It was an accurate price for the old way of working. What changed was not anyone's competence. It was the method. The bottleneck moved from building the thing to specifying and checking it, and those turned out to be things one person with deep context could do quickly. The estimate did not lag reality because people were wrong. It lagged because the org was still pricing the work the way it always had.
Two Kinds of Engineer
Watching this land on teams, I have seen the same split more than once. Some people get defensive. The work touches their territory, and the instinct is to guard it, to slow it, to find the reason it does not really count. Others do the opposite. They see the mechanical part leaving and they move toward the judgment, the domain, the customer, the parts that were always the point. The best engineers I know are becoming more product-like, not less valuable, because they are filling the gap the agent cannot.
Which response you have is turning into a real fork. The capability rewards the people who adapt to it and quietly strands the people who defend against it, and no title protects you from that either way.
Structure Lags Capability
The deepest version of the lag is not any one person. It is the chart itself. An organization is a frozen answer to the question of how work should be divided, and it keeps giving that answer long after the question has changed. Teams do what was always done instead of what is now possible, because the boundaries, the headcount, and the incentives were all set for the old distribution of effort.
That is the real bottleneck now, and it is a more comfortable one to sit in than to name. The tools are ready before the org is. The people who get the most out of this moment are the ones who act on the new reality a little before the structure around them catches up to it, and who are patient with the fact that it will catch up slowly.