Skip to navigation

How NVIDIA NeMo Fabric Fits Within the Agent Ecosystem

View as MarkdownOpen in Claude

The agent ecosystem includes applications, evaluation systems, rollout platforms, and the harnesses that execute agent tasks. NVIDIA NeMo Fabric connects these systems through a common agent execution interface, so they can run supported harnesses and custom agents without implementing each harness integration.

For NeMo Fabric’s capabilities, refer to Overview and Key Features. For its internal design and execution boundaries, refer to Architecture Overview.

The following diagram shows how external projects use NeMo Fabric to run agent harnesses, shared frameworks, custom agents, and remote services and collect normalized execution evidence.

NeMo Helix, NemoClaw, NeMo Gym, Polar, and Harbor use NeMo Fabric's common execution interface to run adapter targets and collect results, artifacts, errors, events, and telemetry references.

NeMo Fabric supplies the agent execution interface. The related products below own application deployment, task provisioning, evaluation, or training workflows. Explore the following projects to understand their roles and find their upstream integrations.

Deployment and Applications

These projects use NeMo Fabric within agent deployment and management workflows:

ProjectPurposeNeMo Fabric’s Role
NVIDIA NemoClawAn open-source reference stack for running supported AI agents inside NVIDIA OpenShell sandboxes. Its v1 development branch uses desired-state configuration and coordinates sandbox provisioning, inference routing, and deployment lifecycle.Manages agent runtimes across supported harnesses. Select a harness through configuration while keeping the same plan, apply, and management workflow. Refer to NeMo Fabric harnesses.
NVIDIA NeMo HelixAn open-source control plane for improving and hardening production agents. Manages deployment, model infrastructure, and evaluation workflows.Translates platform-managed agent definitions into NeMo Fabric configuration and runs the selected harness. Refer to Agent management and Evaluate with a NeMo Fabric harness.

Benchmarking, Evaluation, and Training

These projects use NeMo Fabric to run agents while retaining their own task, evaluation, and training workflows:

ProjectPurposeNeMo Fabric’s Role
Harbor FrameworkEvaluates and optimizes agents and language models. Owns task provisioning, environments, verification, and scoring.The FabricAgent integration runs supported harnesses through NeMo Fabric. Use the same benchmark tasks to compare harnesses, models, and agent configurations.
NVIDIA NeMo GymEvaluates and improves models and agents using task environments. Owns tasks, environments, verifiers, and training workflows.Runs interchangeable harnesses through the NeMo Fabric agent. Compare harnesses and configurations in the same environments, or collect rollouts for training.
NVIDIA NeMo PolarA reinforcement learning rollout framework for real-world agent harnesses. Manages runtime preparation, model routing, trajectory capture, and evaluation.The NeMo Fabric preset provides one entry point for selecting and running supported harnesses. Switch harnesses through configuration while keeping the same task and rollout workflow. Refer to the research paper.

Integration Status

The following table summarizes each project’s integration status and links to its upstream implementation or documentation.

ProjectIntegration StatusUpstream Reference
NVIDIA NemoClawAvailable on the v1 development branchNeMo Fabric harnesses
Harbor FrameworkAvailable upstream; current main targets NeMo Fabric 0.6.xFabricAgent
NVIDIA NeMo GymAvailable upstreamNeMo Fabric agent
NVIDIA NeMo HelixAvailable platform-managed workflowsAgent management
NVIDIA NeMo PolarAvailable on the upstream stable branchNeMo Fabric preset

Build an Integration

Use the NeMo Fabric Python SDK to configure agents, manage runtimes, and consume normalized results and execution artifacts.

Consumer integration guide · Python SDK guide

Roadmap

Potential integrations to explore: