Ask most founders in the AI boom what they are building and you will hear about capability. Faster models. Smarter agents. Longer context windows. Vin Sharma spent three decades inside that machine, and he came away convinced the industry was measuring the wrong thing. The problem was not that AI agents were not smart enough. It was that nobody could bring themselves to trust them.
Roughly 95 percent of enterprise AI projects never reach production. Sharma likes to cite that figure because it reframes the whole conversation. A model that dazzles in a demo and then stalls before deployment is not a technical failure - it is a trust failure. The people who would have to answer for a hallucination, a leaked record, or a prompt-injection attack simply cannot sign off. So the project dies quietly, somewhere between the proof of concept and the pager rotation.
Vijil, the company he co-founded in 2023 with fellow AWS AI leaders, exists to close that gap. Its pitch is deceptively plain: treat trust as infrastructure, the same way you treat storage or networking. Give developers hardened components to build agents with, tools to test reliability and security during development, ways to mitigate risk before deployment, governance at runtime, and a feedback loop that keeps learning from what happens in production. Do that, the argument goes, and the 95 percent starts to move.
Sharma frames the whole thing in human terms. "We cannot trust autonomous agents today, no matter how intelligent they may seem, the way we trust the people we employ," he has said. His point is that human trust was never a feeling - it was a system, built up over millennia through reputation, verification, and consequences. Agents, he argues, need their own version of that machinery before anyone should hand them the keys.
It is a thesis that only makes sense coming from someone who has watched the field from the inside for a very long time.