Vinit Agrawal, co-founder and CTO at Tars, builds this around one line on an enterprise RFP: how do you verify the accuracy of responses by your AI agents? For fifty years that question had a boring answer. Software was rules, rules could be tested, and a green test suite meant the thing worked. An LLM based agent does not work that way. It predicts rather than follows instructions, and that is a property of anything that speaks a human language, not a flaw waiting to be patched in a later model.

The argument runs on an invented case, two companies he calls A and B, both shipping the same kind of e-commerce support agent. The incident that starts it is a support agent that promised a refund to a shopper who was not eligible for one. Nobody wrote that instruction. There is no line of code to point at and no bug to fix. One company treats it as a bug, edits the system prompt, tries a dozen conversations, and tells the client it is resolved. The agent stops promising refunds and starts refusing them to shoppers who are entitled, and nobody notices, because nobody is looking. The other company reads and hand-labels hundreds of real conversations first, then automates the reading. A second model scores every conversation the way a support manager reviews a rep, and that scorer is tuned against the hand-labeled set until it agrees with human reviewers often enough to be trusted on the thousands nobody will ever read. Without that hand-labeled set you would just be asking who checks the checker, forever.

Watch this if you have to answer for an agent's accuracy in front of a client, a regulator, or a board. The closing point is the one to carry into your next vendor call: everyone has access to the same models now, so the advantage is knowing exactly how reliable yours is when nobody else can say.