Score your own agent on the 25-point PROOF rubric — five categories, checkable sub-criteria, no gut calls. Answer honestly; the rubric doesn't care how good your demo looks.
Built from Vishal Mysore's PROOF essay (dev.to) — Planning · Reasoning · Orchestration · Observability · Feedback. All content glory to the original poster.
Vishal scored the same job done two ways — in healthcare and in manufacturing — and the pattern held both times:
A single unverified LLM call with real-world consequences riding on it. A live model wired straight to a physical actuator. A hazard, not an agent.
Same job, same model possibly — but every consequential action grounded in a specific checkable piece of evidence, with a human or a rule-based check signing off before it acted.
His punchline: the gap isn't model quality. Don't spend your review time on whether the plan looks sophisticated — ask to see the trace of the last high-stakes decision, and who had to sign off before it acted. Read the full essay →