Enterprises are not failing AI agents because they cannot do the work. They are failing because agents do it inconsistently, unpredictably, and sometimes unsafely once real-world conditions shift.

That is the core message from Bryan Silverthorn, Director of AGI Autonomy at Amazon, speaking at VB Transform 2026 on Tuesday. Cisco data shows 85% of enterprises are piloting AI agents, but only 5% have shipped them to production. Silverthorn argued the gap is not fixed by better benchmarks alone. In his view, reliability is the blocking problem, and it is tangled up in ways most evaluations still miss.

To make that concrete, Silverthorn framed reliability as four distinct dimensions: consistency, robustness, predictability, and safety. He credited this breakdown to research from Princeton. The point is simple, but sharp: internal evals often entangle these factors, so an agent can look great on a test while still failing in deployment. “It unpacks different factors that I see tangled together in almost every eval I’ve ever seen,” he said. In other words, teams are measuring one thing, but customers are experiencing another.

The “passed internal evals, failed real customers” pattern has a predictable smell in the wild. Silverthorn described a customer deploying an agent for software QA, focused on serial number extraction from screens. For two months, it worked flawlessly. Then it started intermittently reading wrong numbers. The culprit was not some obvious model collapse. The underlying vision encoder behaved differently depending on where the serial number appeared on the screen. Then a software change that humans could not perceive triggered the failure. The lesson, Silverthorn said, is about measurement, not just models. Even when models improve, if your evaluation does not map variability to the stakes of your application, production will eventually expose the blind spots.

This is where a lot of enterprises get stuck in the “pilot purgatory” loop. VentureBeat presented proprietary research before the session reinforcing the same failure mode: half of surveyed companies shipped agents that passed internal evals but failed real customers. And enterprises overwhelmingly track uptime while ignoring accuracy, which is like checking the pulse without checking the diagnosis. Silverthorn’s framework explains why. Uptime tells you the system stayed alive. It does not tell you whether it remained correct, stable, and safe under the messy conditions that matter. Enterprises, in his telling, often default to the model makers’ own evaluations and little else, leaving their testing strategy “a coin flip” between trusting the vendor and trusting nothing.

Silverthorn’s most memorable prescription was cultural, not technical, and it comes from Amazon’s own operational approach inside its AGI lab. Researchers there literally call their agents “interns,” as in, “I'll have my intern talk to your intern.” The joke signals an important mindset: agents can be powerful and occasionally clueless. Managing them, he argued, requires management skills rather than software skills. That means asking what could go wrong, building backups and undo capabilities, and deciding what level of risk is actually acceptable. “You can ask the intern, 'Hey, what might you do wrong here? How might you mitigate your negative outcomes?'” he said.

Amazon’s lab has embraced trade-offs that would make many enterprise risk committees nervous, at least at first glance. It accepts that agents will occasionally run the wrong experiment in exchange for research velocity, including one agent running experiments around the clock on its own high-level research plan. The second-order implication for enterprise leaders is that you do not just need guardrails. You need an operating model for learning while limiting harm. That includes measurable reliability targets tied to what the business is willing to tolerate when things drift.

Silverthorn also set boundaries around what is and is not ready today. Self-improving AI is still “a loaded term.” He said Amazon uses AI to improve its models constantly, but fully autonomous self-improvement remains distant. Computer use is also a core focus for his lab, with a commercial trucking customer using browser automation to stitch together warranty claims across fragmented systems. But he stressed no future agent will rely on computer use alone. Agents should work alongside MCP, APIs, and other tools to complete end-to-end workflows. LLM-as-judge techniques are promising, but they are not the only strategy for aligning agent capability with acceptable risk.

For decision-makers staring at the 85% pilot rate and the 5% production rate, the strategy shift is blunt. Stop asking whether your agent can do something impressive once. Start asking whether it can do it correctly a thousand times in a row. Silverthorn’s underlying wager is that the enterprises that escape the ceiling will not necessarily have the smartest agents. They will have the best managers: teams that decompose reliability, test variability realistically, and run agents with explicit risk controls. In a world where regulators increasingly care about operational behavior, not just model demos, that management rigor may matter as much as the algorithm itself.