Deploy Self-Evolving Agents for Faster, More Secure Research with a Hermes Agent and NVIDIA NemoClaw

NVIDIA Details Self-Evolving Agent for Secure Enterprise Research

NVIDIA has released a new open-source example for building self-evolving AI agents that can securely access both internal data sources like Outlook and Slack and public data from GitHub. The approach addresses a significant enterprise challenge by demonstrating a practical method for creating agents that improve with use without compromising sensitive corporate information. By combining the Hermes Agent with the NVIDIA NemoClaw framework, the system can learn new skills and reporting formats directly from user interactions, making its capabilities persistent across deployments.

The Technical Architecture

The solution is built on a three-part stack designed for secure and adaptive operation. The agent's ability to learn is managed by Hermes Agent, which runs within a secure runtime provided by NVIDIA OpenShell. OpenShell enforces strict network policies and manages credentials to prevent data exfiltration. The entire system is built on the NVIDIA NemoClaw blueprint, which uses models like Nemotron 3 Super for reasoning. This architecture ensures that even if an agent were compromised, it could not send sensitive internal data to unauthorized external endpoints.

• Model: NVIDIA Nemotron 3 Super provides the core reasoning and tool selection capabilities.
• Harness: Hermes Agent manages skills, memory, session state, and integration with data sources.
• Runtime: NVIDIA OpenShell provides a secure sandbox, credential brokering, and network policy enforcement.

The significance of this release extends beyond its immediate use case for research triage. The architectural pattern of a sandboxed, self-improving agent that safely combines public and private data is applicable to a wide range of business functions, including sales intelligence, customer support, and engineering workflows. By providing a full-stack, open-source reference implementation, NVIDIA is outlining a foundational framework for enterprises to move from static, proof-of-concept agents to persistent, production-ready systems that learn and adapt over time. Observability is also built-in, with support for Agent Trajectory Format (ATF) traces that can be streamed to debugging tools like Arize Phoenix.

Strategic Takeaway: NVIDIA's release isn't just about another agent; it's a strategic move to define the enterprise agent stack. By providing an open-source, secure, and self-improving blueprint with NemoClaw and OpenShell, the company is positioning itself as the core infrastructure provider for production-grade agentic AI, moving the conversation from model capabilities to secure, persistent deployment.