An agent needs a job, not a persona

Give a system a role, a few tools, and a manner of speaking, and you have a demo. Give it a job and you have a chance of fitting it into real work.

A job has inputs, a definition of done, a limit on steps, a constrained set of actions, and a person who owns the outcome. Persona is how the system talks. The job is what it is allowed to finish. I keep seeing those swapped. The persona gets the attention because it is the part you can hear. The job is the part that decides whether anyone should let the system touch the work.

Agents disappoint, in the cases I care about, when nobody decided what finished looks like, or who answers when finished is wrong. Character is the easy part to add. Accountability is the part that makes the system usable. Once the job repeats, it should start looking like software: a workflow, a few narrow decisions, and a person on the branches that are still open.

The risk in this view is a job specified so tightly that the system cannot touch the case that is actually new. That case still wants a person, or a model allowed to explore. The job is the boundary around the routine. It is not a claim that every task is routine.