Designing an agent's role around the task

UPMI's published report focuses on task authority and workflow design rather than the number of AI tools in use. Saying that an agent writes proposals does not explain which sources it can read or what it may change.

Defining the inputs, deliverable, and completion standard gives the operator and agent a shared task. A clear scope also makes failures easier to investigate.

A proposal workflow can be divided into drafting, review, and delivery. Drafting creates a structure from a defined source set; review checks figures and evidence; delivery confirms the recipient and final version.

This is an illustrative workflow. It is not a claim that a particular client uses this setup or that every stage has been automated.

Recording the source material and the reasons for human revisions helps reduce repeated mistakes. Connecting the source version to the output version is part of that process.

Measures can include unsupported claims, passages needing revision, repeated work, and review waiting time as well as drafting speed. If a faster draft takes longer to check, that cost belongs in the assessment.

Before your next agent task, write down which inputs it may use, how you will verify the result, and where it should stop if evidence is missing.

These questions apply to research reports, teaching materials, and personal server operations. The tools differ, but the need for a verifiable deliverable remains.