Selected, not hired: the agent lifecycle in five verbs

Organisations reach for the language of employment when they talk about agents because it is the only workforce vocabulary they have. The vocabulary is useful as a prompt and dangerous as a template. Here is the lifecycle in five verbs, each paired with the verb it replaces.

Selected, not hired

A person is hired on the strength of experience, potential and fit. An agent is selected the way a supplier is qualified, against three things no procurement process currently asks: the task family it will work in, defined by the work rather than the software category; the verification cost of that family, in human minutes per output; and the named human role that will carry accountability for what it produces. No signature on the third line, no selection, however capable the agent. This rule alone would have prevented most of the agent failures reported so far.

Deployed, not onboarded

People are onboarded into a culture. Agents are deployed into a boundary. The deployment record states, in the language of the task rather than the model, what the agent may do without review, what it may do only with review, and what it may never do. The record is owned by the accountable human, not by IT, because it is a statement about the work, not about the system. A boundary document does for an agent what a culture does for a person: it tells it where the edges are.

Evaluated, not reviewed

A performance review is a conversation about growth. An agent's evaluation is a measurement, on four scales: output volume; output quality judged by the standard applied to a person doing the same task; incident rate; and the human verification time the agent consumes. The fourth decides whether the agent is productive at all, and it is the one most dashboards omit. An organisation that finds itself asking whether an agent is engaged has lost the thread.

Updated, not developed

People develop through training and experience. Agents change when their vendors release new versions, which means an organisation's artificial workforce can change overnight without anyone deciding it should. A model update is a redeployment: the boundary is re-tested, the evaluation is re-run, the accountable human signs again. The alternative is learning, in an incident review, that the agent that failed is not the agent that was approved.

Retired, not terminated

Agents are retired when the task family changes, when a better agent is selected, or when evaluation falls below standard. Retirement raises a question employment never had to ask, because people carry their skills with them when they leave: does the organisation still have anyone who can do this work by hand? If the answer is no, the retirement is not a decision; it is a dependency revealed.

Who runs these five stages? Each one crosses a boundary between the people and the systems, between the business unit and the whole, between this quarter and year five. No existing function spans all of those. The vendor runs the model. IT runs the platform. The business unit runs the task. Nobody runs the lifecycle, and the lifecycle is where the risk lives. That is the operational core of the case for a Chief Workforce Officer: not a philosophy about the future of work, but five verbs that currently have no subject.