Having spent over a decade at Amazon, and earlier years in regulated, compliance-heavy environments, it's not a surprise that AWS gave its Summit keynote to its VP of Agentic AI and spent it on governing AI agents. Not just a product launch. A lagging indicator. As agents take on more human work (gathering context, making decisions, executing), we've outrun our ability to hold them accountable or verify they can be trusted. So Amazon is leaning hard into trust.
No surprise to anyone who's been inside, and the right move as accountability gets real. This isn't hypothetical. A federal class action alleges UnitedHealth used an AI tool to deny care under Medicare Advantage plans, with about 90% of appealed denials ultimately reversed. UnitedHealth disputes that characterization. Whatever a court decides, the lesson holds: when an agent drifts past its objective, the company doesn't get to disown the result. You can't trust and not verify.
That isn't strategy, it's misplaced hope. Trust is hard-earned and easily lost. So Amazon is doing the right thing. These tools are needed to run agents at scale. A year ago, most of this was duct tape and hope. Now the biggest cloud on earth is treating agent governance as first-class: isolated execution, red-teaming, controls that scale. That's real progress. But ... it's not enough. These controls govern what the platform can see, and they're optimized for Amazon.
Great if you live entirely on AWS, but most of us don't. Agents run across harnesses, frameworks, and models (Gemini, DeepSeek, the rest). A single vendor's control plane isn't enough for a multi-model, multi-cloud reality. One of my favourite analogies is the old streetlight story, where someone is searching for keys under the lamp because it's where they can see. AWS is lighting up the infrastructure. But the keys are mostly out in the dark, in what the agent decides and does.
Agents are becoming behavioural. They take human-like actions. So "trust and verify" can't only mean watching infrastructure. It means verifying behaviour, policy, business rules, and mapping that back to the frameworks you answer to. That part was never AWS's to own. So the takeaway can't be "AWS has governance covered now, relax." AWS would say the same. This is one slice, not the whole story. A company doesn't run one agent, it runs hundreds, and soon thousands, all making decisions.
You need to see across them first, enforce your own rules, then link evidence to the frameworks you're held to. That's what we set out to build at Gentrail months ago. Technical guardrails are one piece of it. Skip the platform-versus-no-platform debate and ask one thing of every agent in production: what was it allowed to do, and what did it actually do? If you can't answer that in a few minutes, that's your gap, and no platform default will close it for you.
For a lot of teams the honest answer is still "not yet." That's not a failure. It's a reflection of where the real work is now.


