Architecture: An Adaptive Agent Stack for Hermes Agents

The NVIDIA NemoClaw open source blueprint deploys Hermes agents with the runtime, models, tools, and controls needed for enterprise use. The architecture gives Hermes a governed execution environment where agents can reason over approved context, call approved tools, and keep running across long-lived team workflows.

With NemoClaw, teams can run Hermes locally, in the cloud, or on enterprise hardware while keeping control over which systems, data sources, and actions the agent can access.

NemoClaw installs the components needed to run Hermes Agents: the NVIDIA OpenShell runtime for policy-based execution, model integrations such as NVIDIA Nemotron, approved tool and data connections, and the persistent Hermes Skills layer.

Core Architecture

• Hermes agent layer: Handles team requests, coordinates planning and reasoning, and returns task-ready responses in shared workspaces.
• Governed runtime: Executes agent actions with configurable privacy, security, and policy controls.
• Model layer: Uses NVIDIA Nemotron and compatible open models for reasoning, summarization, and task automation.
• Approved tools and data: Connects Hermes to systems such as code repositories, documentation, tickets, logs, and internal services.
• Hermes Skills layer: Stores repeatable workflows, preferred formats, tool patterns, decision rules, and expert guidance so future requests improve.
• Deployment layer: Supports local, cloud, or enterprise hardware deployment based on team and IT requirements.

How the Blueprint Works

A user or team asks Hermes to complete a task in a shared environment. Hermes reasons over the request, checks available context, and uses approved tools through the governed runtime. Team members can correct or refine the response, and repeated guidance is captured as Hermes Skills. Over time, those skills help Hermes apply the team's preferred way of working to future requests and can support a governed Hermes Collective Wisdom layer across the organization.