AWS previews Bedrock Managed Agents built with OpenAI, running the Agents API under native IAM

Bedrock Managed Agents is the most structural part of AWS's growing OpenAI partnership. Instead of only hosting OpenAI models, AWS is running an AWS-native version of the OpenAI Agents API itself. The managed runtime offers:

- multi-step tool orchestration

- persistent sessions

- sandboxed code execution

- support for Model Context Protocol (MCP) servers

All of it runs under standard AWS IAM governance. The preview is available in three US regions.

For enterprises, the appeal is governance. Teams that have built on OpenAI's agent abstractions can now run those agents inside their existing AWS security boundary. IAM roles, logging and network controls apply as they do for other AWS services. The AWS Builder Center's fall 2026 update presents this as the centerpiece of AWS's agentic strategy, alongside new builder tools.

The competitive picture has shifted. The same week, Bedrock added GPT-6 Sol and Luna, GPT-6.1 Sol, Claude Opus 5.5 and Grok 4.7. Bedrock is positioning itself as a neutral model marketplace and, increasingly, a managed agent runtime. That puts AWS head-to-head with Microsoft Foundry, Google's Vertex AI agents and Anthropic's Claude Managed Agents. AWS is also building its own governance layer: it released the open-source Dogwood policy language and a local engine to enforce it.

Caveats:

- This is a preview in only three US regions.

- Pricing and quotas were not detailed in the sources.

- Security posture is a live question in a week dominated by the Hugging Face agent breach. The sandboxed code execution claim will be scrutinized, and builders will want to know whether NVIDIA's OpenShell or Sentry layers can sit underneath.

The broader signal is that OpenAI is distributing its agent stack through a hyperscaler other than Microsoft. That sharpens questions about how exclusive the Azure relationship still is.