New updates on AWS GovCloud (US) | Issue #7 (2026)
August 16-31, 2026 | New updates on AWS GovCloud (US): AI Without Compromise. Choice Without Complexity. The model portfolio on Amazon Bedrock in AWS GovCloud (US) now spans seven families: Amazon Nova, Anthropic Claude, Google Gemma, Meta Llama, NVIDIA Nemotron, OpenAI GPT, and xAI Grok. This issue covers what the expanding lineup means for government AI strategy and how AgentCore’s latest capabilities, Memory, Policy, and Managed Harness, give teams the tools to deploy agents at mission speed. More models. Same boundary. Built in.
Executive Insights
AI Model Choice Brings Mission Advantage
Your mission now picks the AI model. AWS GovCloud (US) now offers multiple AI model families through Amazon Bedrock, including Amazon Nova, Anthropic Claude, Google Gemma, Meta Llama, NVIDIA Nemotron, OpenAI GPT, and xAI Grok, all inside the compliance boundary. As AWS CEO Matt Garman has shared, customers should never be locked into a single model. For government missions, model choice is how agencies stay in control of their technology systems, promote innovation, and deliver outcomes at the pace the mission demands. Model choice directly supports America’s AI Action Plan, which calls on government to accelerate AI adoption and promote innovation across federal missions. AWS’s up to $50 billion infrastructure commitment underpins these models with the capacity to run them at the scale the mission demands. A mission system can use a lightweight model to classify sensor data, a reasoning model to generate threat assessments, and a code model to automate patching, all through a single access point with no additional procurement or vendor onboarding. Read more on the AWS Public Sector Blog.
Solutions Insights
Amazon Bedrock AgentCore – Memory, Policy, and Managed Harness in AWS GovCloud (US-West)
Here's what practitioners need to know about the new Amazon Bedrock AgentCore capabilities now available in AWS GovCloud (US).
Amazon Bedrock AgentCore, the fully managed platform for building, connecting, and optimizing AI agents, has expanded its feature set in AWS GovCloud (US-West) with three new capabilities: Memory, Policy, and Managed Harness. Building on AgentCore Runtime's existing FedRAMP Class D (formerly High) authorization, these three capabilities expand the platform for production agent deployment in regulated environments. Teams can now build context-aware, policy-governed agents and move from prototype to production faster with the organizational controls required for regulated workloads. All data is processed entirely within the isolated AWS GovCloud (US-West) Region.
Memory gives agents both short-term and long-term recall without requiring teams to manage complex memory infrastructure. Short-term memory captures immediate conversation context within a session, enabling coherent multi-turn interactions. Long-term memory automatically extracts persistent insights and user preferences across sessions, so agents become more intelligent and personalized over time. For government teams, this means an agent assisting with procurement workflows can remember a user's preferred contract vehicles, past vendor evaluations, and recurring compliance requirements across sessions, eliminating repetitive context-setting while maintaining data residency within the AWS GovCloud (US) boundary. Long-term memory is always encrypted at rest using AWS KMS. By default, encryption uses an AWS-owned key. For additional control, teams can optionally configure a customer-managed KMS key.
Policy provides centralized, fine-grained controls for agent-tool interactions through natural language policies. Teams author rules in plain English (e.g., "Allow the grants agent to query funding databases only for the user's assigned program office"). These natural language policies are automatically transpiled to Cedar, AWS's open-source authorization policy language, and formally verified before enforcement — ensuring policies behave exactly as intended before execution. Policies attach to an AgentCore Gateway that evaluates every tool-access request before allowing or denying it. Cedar enforces default-deny semantics: no policy means no access. This architecture separates authorization logic from agent code, giving security teams auditable, centralized control over what agents can and cannot do — a critical requirement for FedRAMP Class D environments. Policy evaluation metrics, including allow/deny decision counts and determining policy identifiers, are published to CloudWatch by default. For detailed per-request audit trails (authorization decisions, reasons, and tool-level allow/deny lists), enable CloudTrail data events for Gateway resources and activate Gateway traces.
Managed Harness eliminates the need to write orchestration code. Developers declare an agent's model, tools, and instructions through configuration and deploy it with minimal API calls. AgentCore handles environment, compute, memory, identity, and observability management. The harness is model-agnostic, meaning teams can switch foundation model providers without losing session context or rewriting infrastructure. From a single configuration definition, a production-grade agent runs in its own isolated environment with filesystem, shell access, memory persistence, and web browsing. When custom orchestration is needed, a single CLI command exports to Strands-based code for full flexibility.
Together, these three capabilities address the core challenge regulated teams face when operationalizing AI agents: maintaining security and compliance controls while moving at the speed mission demands. Memory delivers personalization without data leakage, Policy enforces least-privilege tool access with full auditability, and Harness removes infrastructure overhead so teams focus on mission logic rather than plumbing.
To learn more, visit the Amazon Bedrock AgentCore documentation and the AWS Public Sector blog: Deploy AI agents in AWS GovCloud (US) using Amazon Bedrock AgentCore.
Service and Feature Releases
Source: What's New Posts, 08/16/26 - 08/31/26.
Previous AWS GovCloud (US) Newsletters
Helpful Links
Did you find what you were looking for today?
Let us know so we can improve the quality of the content on our pages