Work is moving into agents. Every post in this series so far has said some version of that. What nobody wants to say plainly is what happens to the people whose job used to be doing that work by hand.
Two answers get repeated constantly, and both are wrong. One says nothing changes, agents just make everyone a little faster. The other says the roles disappear outright, down to a skeleton crew. Neither matches what I have watched happen on my own team over the last year.
Where the job goes instead
The role does not vanish, it moves. Some of what a job used to be gets picked up by an agent, and what is left concentrates into a smaller, sharper set of responsibilities only a person can still do. Pretending the move is not happening, or pretending it erases the person on the other end of it, both dodge the real work: figuring out what that smaller job looks like.
I have watched this land differently across four roles I work with every day.
Four roles, four moves
1/ Product managers. The old job was speccing screens, writing a PRD for what a human would click and see. That goes away first. The new job is writing capability contracts, what the agent can and cannot do, and eval definitions, how you know it did the job right. The agent is a persona now, not a UI. A PM who cannot write an eval is roughly where a PM who could not write a spec was five years ago.
2/ Designers. The old craft was a twenty-screen flow, every state and edge case drawn out by hand. Fewer of those flows need to exist once an agent handles the middle of the task. What does not go away is the checkpoint, the moment a person looks at what the agent produced and decides whether to trust it. Designing that moment, what it shows, how it earns trust instead of assuming it, is the harder design problem now. The review surface is the new craft, not the twenty screens.
3/ Frontend engineers. Fewer destination apps get built from scratch. More of the work is a tool surface, a thin client over an SDK, something an agent calls into rather than a page a person navigates. The skill has not disappeared, it has moved closer to the agent layer, defining the contract the agent runs on instead of the page a human clicks through.
4/ Ops and support. The old job was doing the task, ticket by ticket, call by call. The new job is supervising a fleet of agents doing the task and stepping in on the exceptions. That sounds like less work. It is not. Judgment about when to trust the fleet's output and when not to is a harder skill than the ticket queue ever demanded.
The roles that gain value everywhere
Underneath all four of those moves is the same pattern. Three kinds of ownership get more valuable no matter which of the four roles you started in.
Whoever writes the tool contract, the actual definition of what the agent can call and what it gets back, shapes how the whole system behaves. Whoever owns the evals, the actual test of whether the output is good enough to ship, controls quality more than anyone reviewing screens by hand ever did. Whoever owns a workflow end to end, one person accountable for the outcome instead of a relay of handoffs, is worth more than a team each owning one artifact in the middle of it.
Artifacts lose value in this shift. The screen, the PRD, the ticket closed, all of it used to be proof of work. None of it is proof of outcome anymore, an agent can produce ten times the artifacts a person could and prove nothing about whether the outcome is any good. Outcomes hold their value. They are harder to fake and harder to automate away, because someone still has to decide what a good outcome even is.
What happened on my own team
A VP on my team left this year and I did not backfill the role. That decision came out of real pressure, not a strategy slide. I ran the PM function smaller than it had been, leaned harder on the automation we already had for reporting, first-pass drafting, and status work that used to eat a person's full week, and put the people who remained onto owning outcomes instead of producing artifacts.
We shipped more products than before, not fewer. I want to be precise about what that does and does not prove. It does not prove people are unnecessary, someone still had to decide what to automate and own the evals and be accountable when an agent's output was wrong. What it proves is narrower: a smaller function, built around outcome ownership, can out-ship a larger one built the old way.
A decision like that puts real weight on whoever is left to prove the function still works. The folks who came through it more valuable, not less, were not the ones who had produced the most decks or specs under the old structure. They already owned something end to end, a workflow, a relationship, an eval, and could point at it and say that is mine.
Planning your own next two years
If you are planning the next two years of your own role, here is what I would do with this, an operator's checklist, not a research finding.
1/ Stop measuring your value by artifact volume. If your update to your manager is a list of screens shipped or tickets closed, start a second list next to it, the outcomes you personally own start to finish.
2/ Find the eval nobody owns yet. Every agent workflow running near you has a quality bar somewhere, defined or not. Go define it. That ownership is durable in a way a single feature never was.
3/ Write the tool contract for one workflow you touch, even an informal one: what goes in, what comes out, what counts as correct. Whoever writes that document tends to end up running the workflow, plan or not.
4/ Do not wait for a reorg to tell you your role changed. By the time it is official, the people who already moved are the ones the reorg gets written around.
None of this is about whether agents are good or bad for the people doing this work today. That debate misses the point. The work moves, the org chart eventually follows, and the people who own outcomes instead of artifacts are the ones still standing in whatever shape the chart takes next.
This is post 8 of 9 in my Thoughtful interfaces series. The rest lives at hagestedt.com/writing.