[AWS describes an enterprise agent-access pattern](https://aws.amazon.com/blogs/machine-learning/govern-ai-agent-access-with-okta-cross-app-access-and-amazon-bedrock-agentcore-part-2/) that combines an AgentCore Gateway entry point, Okta grants, Cedar policies binding tools to agents, and traces of access approvals and denials. This is an implementation example, not proof that every deployment using these products is governed.
The executive implication: evaluate authority alongside capability. Before approving another agent pilot, require a named business owner, permitted tools, user-authority boundaries, access logs, and a tested revocation procedure. Ask for a demonstration of a denied action—not just a successful one.
[AWS’s Taxbit Assist case study](https://aws.amazon.com/blogs/machine-learning/taxbit-assist-a-grounded-role-aware-ai-agent-on-amazon-bedrock-agentcore/) describes an assistant grounded in curated knowledge, with citations, role-aware controls, and ongoing production answer-quality measurement. It is not a dental deployment, but it offers a useful design reference for a DSO pilot.
For a DSO, consider starting with approved operating policies rather than clinical recommendations. Limit the pilot to one administrative workflow, define access by role and location, and assign an owner to each source document. Require cited answers and a clear escalation path when the approved material does not resolve the question.
[CISA added CVE-2026-19868 to its Known Exploited Vulnerabilities Catalog](https://www.cisa.gov/news-events/alerts/2026/11/06/cisa-adds-one-known-exploited-vulnerability-catalog), citing evidence of active exploitation. The record identifies a SonicWall SonicOS SSLVPN use-of-single-factor-authentication vulnerability.
Treat this as an exposure-management priority, not simply another advisory to circulate. Have security and infrastructure teams confirm whether affected technology is present, identify externally reachable instances, and document remediation or containment. Ask for evidence of closure and explicit ownership of any exception. Do not let the AI agenda displace this work.
[AWS’s Amazon Quick taxonomy](https://aws.amazon.com/blogs/machine-learning/taxonomy-of-agentic-primitives-in-amazon-quick/) organizes agentic work around triggers, agents, capabilities, and learned context. Use that vocabulary as a design aid—not as a substitute for an operating model.
Require a one-page operating specification for each proposed workflow: what starts it, which agent acts, which capabilities it can invoke, and what context it retains. Add a business owner, exception queue, rollback procedure, and acceptance criteria. CIOs and COOs should approve the workflow together, including who handles the work when the agent cannot.
[AWS outlines an MCP server on ECS Fargate](https://aws.amazon.com/blogs/machine-learning/serverless-mcp-server-for-ai-agent-data-lake-access-on-amazon-ecs-fargate/) that gives agents metadata-driven access to an enterprise data lake. In the described implementation, exposing another dataset requires a JSON file rather than per-dataset integration code.
The leadership implication: make dataset approval a separate decision from technical onboarding. Before exposing a dataset, require an accountable owner, agreed definitions, freshness expectations, permitted uses, and access tests. Start with a small approved collection and verify that agents cannot retrieve data outside their assigned scope.
[AWS describes remote Claude Code sessions on AgentCore](https://aws.amazon.com/blogs/machine-learning/extend-claude-code-to-remote-sessions-on-amazon-bedrock-agentcore-with-security-controls/), with each task running in an isolated session inside the customer’s AWS account. Model calls go through Amazon Bedrock, and the output returns as a GitHub pull request for review and merge.
This offers a practical approval boundary: automate preparation while retaining human release authority. Pilot bounded maintenance tasks, restrict repository and credential access, and require normal testing and review. Measure total time through acceptance—including reviewer effort and rework—rather than counting generated changes as completed work.
[SageMaker Inference now supports NVIDIA Multi-Instance GPU](https://aws.amazon.com/blogs/machine-learning/sagemaker-inference-now-supports-multi-instance-gpu-to-maximize-gpu-utilization-for-model-endpoints/), according to AWS. MIG divides a GPU into hardware-isolated partitions, allowing different models to occupy separate partitions or use a fraction of an accelerator.
Treat partitioning as an engineering option, not an automatic savings claim. Ask the platform team to benchmark representative workloads against current endpoints, including latency, throughput, cost, and operational effort. Standardize only the configurations that meet service requirements, and name an owner for capacity planning and troubleshooting.
- [AWS Machine Learning: Govern AI agent access with Okta Cross App Access and Amazon Bedrock AgentCore – Part 2](https://aws.amazon.com/blogs/machine-learning/govern-ai-agent-access-with-okta-cross-app-access-and-amazon-bedrock-agentcore-part-2/) - [AWS Machine Learning: Taxbit Assist: A grounded, role-aware AI agent on Amazon Bedrock AgentCore](https://aws.amazon.com/blogs/machine-learning/taxbit-assist-a-grounded-role-aware-ai-agent-on-amazon-bedrock-agentcore/) - [CISA Advisories: CISA Adds One Known Exploited Vulnerability to Catalog](https://www.cisa.gov/news-events/alerts/2026/11/06/cisa-adds-one-known-exploited-vulnerability-catalog) - [AWS Machine Learning: Taxonomy of agentic primitives in Amazon Quick](https://aws.amazon.com/blogs/machine-learning/taxonomy-of-agentic-primitives-in-amazon-quick/) - [AWS Machine Learning: Serverless MCP server for AI agent data lake access on Amazon ECS Fargate](https://aws.amazon.com/blogs/machine-learning/serverless-mcp-server-for-ai-agent-data-lake-access-on-amazon-ecs-fargate/) - [AWS Machine Learning: Extend Claude Code to remote sessions on Amazon Bedrock AgentCore with security controls](https://aws.amazon.com/blogs/machine-learning/extend-claude-code-to-remote-sessions-on-amazon-bedrock-agentcore-with-security-controls/) - [AWS Machine Learning: SageMaker Inference now supports Multi-Instance GPU to maximize GPU utilization for model endpoints](https://aws.amazon.com/blogs/machine-learning/sagemaker-inference-now-supports-multi-instance-gpu-to-maximize-gpu-utilization-for-model-endpoints/)
Know an executive who should read it first? Forward this.
— BWP
Tell me what to research next.
Two questions: which topics matter most to you, and what challenges you're trying to resolve right now — including doctor or hygienist turnover. Your answers shape upcoming issues.
Take the surveyWas today's edition worth your five minutes? Your vote shapes what lands in your inbox next.
Know an executive who should read it first? Send it their way.
