OpenAI Managed Agents Arrive on AWS Bedrock, Run Fully Inside AWS

OpenAI and AWS deepened their arrangement at DevDay 2026 with Bedrock Managed Agents: the Managed Agents offering OpenAI has run in limited preview since April now integrates core Agents API features with AWS-side customization and direct connectivity to AWS resources, allowing OpenAI-powered agents to operate entirely within AWS infrastructure. For enterprises that have kept agent projects in pilot because production data cannot leave their cloud perimeter, this is the announcement that changes the answer.

This piece covers what Bedrock Managed Agents actually is, why running agents inside customer infrastructure matters more for agents than it did for plain model inference, who should evaluate it, and what remains unannounced. It is written for platform and infrastructure teams that own enterprise AI deployment decisions.

Bedrock Managed Agents: what was announced

The building blocks are familiar separately: OpenAI models have been available through Amazon Bedrock, and OpenAI's Managed Agents preview has offered hosted agents since April. The new piece is the integration between them — core Agents API features arriving in the Bedrock context, with AWS customization and direct resource connectivity, so the agent runtime itself lives where the customer's workloads already run.

Direct resource connectivity is the operative phrase. An agent that can reach S3 buckets, databases, and internal services natively — under AWS identity and networking controls — rather than through externally exposed APIs is a structurally different deployment than an agent calling in from OpenAI's cloud.

Why agent locality matters more than model locality

For plain inference, data residency is a matter of where prompts and completions travel. Agents raise the stakes: they hold credentials, execute multi-step plans against live systems, and accumulate working state. Every enterprise security review of an agent pilot asks the same questions — what can it reach, under whose identity, observed by whom — and answers improve dramatically when the whole loop runs inside infrastructure the customer already governs with IAM, VPC boundaries, and CloudTrail-style audit.

That is the gap between the 2025-era pattern — agents as SaaS calling webhooks into the enterprise — and what OpenAI is now offering: the agent as a workload in the customer's own cloud, with the same Agents API surface developers use elsewhere. The same DevDay added computer use, Tool Search, and Context Compaction to that API, so the capability set arriving inside AWS is current, not a trailing subset.

Who should care: the enterprise middle

The clearest beneficiaries are organizations with mature AWS estates and stalled agent programs: financial services, healthcare, and any shop whose compliance posture forbids production credentials leaving the boundary. OpenAI's parallel Private Intelligence announcement — zero data retention with private safety processing now, confidential Private Inference this fall — targets the same buyers from the data-handling side; Bedrock Managed Agents targets them from the runtime side.

It also repositions the build-versus-buy debate inside AWS shops. Teams have been assembling agent runtimes from Bedrock primitives and open-source harnesses precisely to keep execution local; a managed OpenAI runtime with the Agents API surface undercuts the main reason to keep maintaining that plumbing — if the feature set and pricing land right.

The competitive frame

For AWS, hosting OpenAI's agent runtime strengthens Bedrock's claim to be the neutral enterprise AI layer, one week after DevDay also showcased OpenAI's Microsoft ties — specialist dots integrating with Microsoft Agent 365. OpenAI courting both clouds' enterprise machinery simultaneously says the agent platform, not any single distribution partner, is the strategic asset.

For competitors, the move raises the bar on enterprise agent deployment: in-VPC execution with a frontier lab's managed runtime is now a checkbox rivals will be asked about in every procurement cycle.

Availability and what to watch

OpenAI described the integration at the recap level; still unpublished are pricing, the precise subset of Agents API features included, region coverage, and the mechanics of resource permissioning — the details that decide real deployments. Managed Agents itself remains rooted in the limited preview that began in April, so expect gated access rather than general availability.

Teams interested should line up the evaluation now: inventory which internal resources an agent would need, map the IAM design for agent identity, and prepare a bounded pilot workload. When access opens, the differentiating test is not whether the agent works — it is whether your security team can audit what it did using the AWS tooling they already trust.