“The largest problem shouldn’t be merely whether or not an agent produces an excellent reply,” he says. “It’s whether or not the group can show what the agent accessed, what directions it adopted, what instruments it invoked, what choices it made, the place a human intervened, and whether or not it stayed inside licensed boundaries.”
With no full stage of runtime visibility, firms are left with screenshots, logs, and after-the-fact explanations that will not meet authorized, compliance, or safety necessities, he says.
Brokers must be constantly verified as an alternative of absolutely trusted, he provides, with governance engineered into the agent structure itself. Governance ought to embody role-based entry, policy-bound execution, human approval thresholds, supply and gear provenance, immutable exercise information, confidence scoring, exception dealing with, and clear escalation paths when an agent reaches the sting of its authority, he recommends.
