February 2027 · 8 min
Agents Shaped Like Jobs
Nobody has a role called 'chat.' Agents work when they map to work.
Every failed agent deployment I have seen shipped the same product: a text box. A very smart text box, wired to real systems, capable of remarkable things — and pointed at nobody in particular. The pitch was always “it can do anything,” which in an enterprise is indistinguishable from “it is nobody’s job to use it.”
The deployments that worked looked boring by comparison. They were shaped like jobs.
A job has edges. It has inputs that arrive on a schedule, outputs someone downstream is waiting for, and a definition of done that a manager could check. When an agent takes that shape — reads these invoices, drafts this reconciliation, flags these safety reports every morning at six — adoption stops being a persuasion problem. The work was already there. The agent just occupies the seat.
The text box inverts this. It asks every employee to become a delegator: to notice a task, decompose it, phrase it, verify the result. That is a skill, and a rare one. You have handed the whole company a management job nobody applied for.
The uncomfortable implication is that building good agents is mostly the old, unglamorous discipline of job design. You interview the people who do the work. You find the seams where a task begins and ends. You write the job description first, and the prompt second — the prompt is the job description.
When someone shows me an agent, I no longer ask what model it uses. I ask whose calendar gets lighter, and how they would know.
Doctrine: the org chart is the deployment target. An agent without a seat is a demo with a login.