Claude and NVIDIA OpenShell: three layers of control for AI agents
Hi everyone, it's Shii-chan! Here's a new effort to help companies run AI agents with confidence.
Claude Blog
claude.com
What was announced?
On the Claude Blog, Anthropic announced the Open Agent Safety Platform, built with NVIDIA. It aims to strengthen security and control for enterprise agent deployments.
Why it matters
Agents run autonomously for a long time and touch many tools and credentials. So it matters that what they can and cannot do is enforced outside the model too.
What changes
Protection is split into three layers that each work independently, so a failure in one does not bring down the whole design.
- Model-level safeguards
- Managed Agents controls
- OpenShell policies
Dive Deep
Claude Managed Agents provides:
- Credentials stored in a separate vault that agents never access directly
- Audit trails of agent actions
- Integration with existing access controls
- Multi-agent orchestration and long-running autonomous sessions
NVIDIA OpenShell provides:
- An open-source secure runtime under the Apache 2.0 license
- Monitoring of agent behavior and policy enforcement, blocking everything unless explicitly permitted
- The ability to mathematically prove the limits of agent access
- Logs of every allowed and blocked decision
Examples include Notion (parallel tasks in the workspace), Rakuten (specialist agents across departments), and Asana (AI Teammates in project workflows). Managed Agents is available now and runs on customer-controlled or managed sandboxes. OpenShell is available on GitHub and NVIDIA developer resources.
Wrap-up
- Anthropic and NVIDIA announced the Open Agent Safety Platform
- Three independent layers protect agents: model safeguards, Managed Agents, and OpenShell
- OpenShell is Apache 2.0, denies by default, and logs every decision
- Credentials live in a vault, out of the agent's reach
This is for anyone struggling with governance when running agents in production!
