Minimum agent operating model:
clear role, one owner, defined scope, forbidden areas, approved tools, checks, escalation path, success metric, stop condition.
If you cannot write that page, the role is not ready.
If an agent needs half the company involved to make one safe change, the prompt is not the problem.
You have an ownership problem, a boundary problem, or an architecture problem.
Usually all three.
Do not start with an AI transformation programme.
Start with one owned role, one clear scope, one before-state measure, and a six-week decision point.
Expand on evidence. Not enthusiasm.
One dependency-upgrade pilot opened 38 PRs in six weeks.
31 merged after minor edits.
Median alert-to-merge time dropped from roughly 3 weeks to under 4 days. Review time stayed flat.
That is the standard. More useful work, same review cost.
One useful agent role I saw did not answer customers.
It routed support escalations with evidence.
214 escalations went through it. 76% were accepted without rerouting. Median time to first technical owner fell from 9 hours to 2.5.
That is what ROI looks like. Constraint relief, not theatre.
If you are measuring agent ROI, start with operational metrics:
lead time, review time, acceptance rate, defect rate, neglected work cleared, full cost of ownership.
Anything else is too easy to game.
PRs opened, lines changed, tickets touched.
Those are activity metrics.
ROI means a real constraint moved without quietly dumping cost into review, rework, or incident risk.