AWS’s new managed agent service, powered by OpenAI, offers a glimpse of how AI vendors could restructure the process enterprises use to start building agents.
The tech giant and AI lab on April 28 unveiled an extended partnership, with OpenAI’s GPT-5.5 and GPT-5.4 models now available in a limited preview on Amazon Bedrock, the tech giant’s generative AI platform. OpenAI’s specialized coding agent, Codex, is now integrated into Bedrock as well. Also, the two vendors introduced a new service called Amazon Bedrock Managed Agents, powered by OpenAI.
The service enables vendors to deploy production-ready OpenAI-powered agents on AWS in a straightforward way, AWS said. The service is built with the OpenAI agent harness, the software infrastructure and control layer that enable a generative AI model to function as an agent. Every agent operates with its own identity and runs in an enterprise environment, with all model inference running on Amazon Bedrock.
The managed service and expanded partnership come a few months after the two vendors signed a multi-year agreement worth $38 billion and a few months after Amazon said it would invest $50 billion in OpenAI. It also comes as agentic AI continues to grow dramatically, with a host of vendors trying to make it easier for enterprises to create and deploy agents.
While many enterprises are dabbling in agentic AI, they have met challenges that could be detrimental to their overall business structure, such as out-of-control agents.
“Increasingly agentic AI is a minefield,” said David Nicholson, an analyst at Futurum Group.
The service is a way for AWS and OpenAI to help enterprises streamline the process of creating and building their agents, said Mark Beccue, an analyst at Omdia, a division of Informa TechTarget.
“It eliminates one step, which is ‘what’s my underlying model?’” Beccue said.
He added that, with Amazon Bedrock Managed Agents, customers no longer have to decide which model to use or how to use it; the choice has already been made, so they do not have to worry about the preferred model when building their agents.
Amazon Bedrock Managed Agents is also a way for enterprises interested in safety and guardrails to experiment with agents, Nicholson said.
“This is going to be another relatively safe haven for people who are being asked to pursue an agentic AI strategy, but who have been confused and concerned up to this point,” he said. “This is more about giving someone a safe and secure option.”
“This sort of collaboration is imperative if any of them wants to get any of the market share in AI from people who are terrified of the negative consequences that we are starting to hear about,” Nicholson continued, referencing a couple of incidents in which agents have gone out of control and taken their own, unauthorized initiative.
A recent instance of this agent misbehavior involved the automotive rental and sales platform PocketoS. The vendor’s Cursor coding agent, powered by Anthropic’s Claude model, reportedly deleted a database and the backup of the database.
“Increasingly, the headlines are going to be horror stories of agents gone wrong, so this helps,” Nicholson said.
He added that OpenAI and AWS will likely not be the only vendors to have partnerships like these.
The type of agent platforms enterprises are drawn to, though, will depend on how much they cost, Beccue said.
“The way they price these things is a deciding factor in whether somebody is going to use it,” he said.
