Amazon Bedrock AgentCore getting key governance updates

When Amazon announced Bedrock AgentCore in July, it felt significant. Here was a major cloud platform adding a component-based way to build agents with many of the pieces a developer would need to feel comfortable building something in an enterprise setting. This included runtime, browser, observability, identity, memory management components, and even a code interpreter.

What was missing, however, was a way to govern agents after they were deployed, something enterprise customers had to be demanding, and an issue Amazon rectified when they announced two new AgentCore pieces at AWS re:Invent this week: policy and evaluation. The former lets companies define exactly what applications, APIs, MCP servers and so forth that an agent is allowed to access and under what conditions. The latter lets developers continually check that the agent is doing what it’s designed to do, even when there is a major change like a model update.

In a blog post, announcing the new updates the company recognized that customers need to be confident when deploying agents inside large organizations and that requires being able to not only observe them, but set up guardrails for their operation to make sure they can’t access sensitive information, any more than certain human employees would be allowed to.

“Development teams must balance enabling agent autonomy while ensuring they operate within acceptable boundaries and with the quality you require to put them in front of customers and employees.” As AWS CEO Matt Garman said in his keynote address this week, it’s essential to set clear boundaries for agent actions, and which tools they are able to access.

Absolutely crucial pieces

When the product was announced last summer, cloud consultant David Linthicum wasn’t terribly impressed, telling FastForward that it appeared to be “a collection of existing AWS tools packaged with some new abstractions,” a position he stands by. 

This week, he added that the governance components were absolutely essential, and was surprised the first version didn’t have them. “I’m not sure how you launch a product like that without governance, but they did,” he said. “Clearly, the market was pointing out the missing pieces, and they are adding them now.”

But Holger Mueller, an analyst at Constellation Research says this just appears to be the kind of update you would expect from V2 of an agent framework. “First you have to build the agents, then you have to govern them,” Mueller said. “As with all governance frameworks, they will need to be battle tested, and we will only know in spring how good they really are.”

While these policy and evaluation pieces are crucial, they still aren’t widely available quite yet. They were announced in preview this week. The company noted that the AgentCore SDK has been downloaded 2 million times since launch. Customers include the PGA Tour and Workday.

Featured image by Ron Miller.