The enterprise AI industry has a math problem. Cisco displays data 85% of enterprises are testing AI agents, but only 5% have put them into production. Horse VB Transform 2026 tuesday, Bryan SilverthorneThe director of AGI Autonomy at Amazon explained why this gap persists and why the answer isn’t better benchmarks.
Silverthorn, who joined Amazon through its acquisition of Adept AI and now leads multimodal agent training in the company’s AGI lab, argued that reliability should be broken down into four distinct dimensions: consistency, robustness, predictability and security — a framework he credits to research from Princeton.
"This opens up the various factors that I see intertwined in almost every assessment I’ve ever seen," he said.
Why AI agents pass internal evaluations but fail real customers in production
The framework is important because agents regularly favor internal assessments and then collapse in the wild. Silverthorn described a customer deploying an agent for QA software that involved extracting the serial number from displays. It worked flawlessly for two months – then it started reading the wrong numbers intermittently. The culprit: the main vision encoder behaved differently depending on where the serial number appeared on the screen, and a software change that was imperceptible to humans caused the failure.
The lesson is about measurement, not just models, Silverthorn said. "Models should be better. Obviously, we are working hard to make the models better," he said. But he added that the deeper implication is that teams need to define metrics of variability and match measurement rigor to the application’s stakes. VentureBeat’s own research presented before the session reinforces this point: half of the companies surveyed send agents who pass internal assessments but fail real clients, and businesses track uptime without regard for accuracy — checking the pulse without checking the diagnosis. A related finding highlighted how few safeguards there are: most businesses abuse model manufacturers’ own assessments and prefer little, leaving a coin flip between trusting the vendor and trusting nothing, as I described the testing strategy on stage.
Within Amazon’s “experiment” framework for driving autonomous AI agents
Silverthorne’s most memorable recipe was cultural rather than technical. At Amazon’s AGI lab, researchers literally summon their own agents "practitioners" – as in "I will share my experience with yours." The joke carries a serious operating philosophy. As practitioners, agents are powerful but sometimes clueless, capable of amazing feats and spectacular derailments.
Managing them requires more management skills than software skills: asking what can go wrong, adding backups and disabling capabilities, and consciously deciding what risk you can accept. "You can ask the intern, ‘Hey, what could you possibly be doing wrong here? How can you reduce your negative consequences?’" he said. Amazon’s lab accepted this trade-off, accepting agents who sometimes ran errands in exchange for research speed, including one agent who ran around-the-clock experiments based on his high-level research plan.
What enterprise leaders need to do before deploying agents at scale
Silverthorn was candid about the limits of today’s technology. Self-improving AI remains "loaded term," he said – Amazon uses artificial intelligence to constantly improve its models, but fully autonomous self-improvement is a long way off. Computer usage remains a key focus of the lab, with a commercial trucking customer using browser automation to consolidate warranty claims across already fragmented systems**, although he emphasized that no future agent will rely solely on computer usage, but will work with MCP, APIs and other tools to complete workflows**. And while LLM techniques as referees are promising, they are only one of several strategies for matching an agent’s capabilities with acceptable risk.
For businesses stuck in pilot purgatory, the way forward starts with a mindset shift: stop asking if your agent can do something impressive once, and start asking if they can do it right a thousand times in a row.
In other words, the businesses that avoid the 85% ceiling will not be the businesses with the smartest agents. They will be the ones with the best managers.
