Skip to navigation

Architecture Overview for NVIDIA NeMo Fabric

View as MarkdownOpen in Claude

NVIDIA NeMo Fabric uses a Python SDK backed by a Rust core to validate configuration and coordinate execution through harness adapters and environment providers. Applications supply tasks and use the results; each harness retains its own agent logic and tools.

The following diagram shows how applications, evaluation systems, and rollout platforms use NeMo Fabric to run harnesses and receive execution evidence.

Applications, evaluation systems, and rollout platforms send FabricConfig through the Python SDK to the Rust core, which normalizes configuration, runs harness adapters, and returns RunResult; optional NeMo Relay captures telemetry.

Components

The components divide responsibility as follows:

ComponentRole in a Run
Application or evaluatorDefines the task, chooses the harness, submits requests, and uses the results. Owns scheduling, concurrency limits, retries, scoring, and recovery.
NeMo FabricValidates configuration and coordinates execution through the harness adapter and environment provider. Returns normalized results and execution evidence.
Environment providerCreates or connects to the environment where the harness runs. Providers such as OpenShell supply sandbox isolation and policy enforcement.
Harness adapterImplements the NeMo Fabric adapter contract for a specific harness. Translates NeMo Fabric configuration and execution requests into native harness operations, then normalizes the results.
Harness or custom agentPerforms the task using its own agent logic, tools, and execution behavior.
NeMo Relay, when enabledInstruments supported harness execution and routes telemetry to configured destinations.

Data Flow

For example, an application asks an agent to review a code change:

  1. Your application submits the task by sending the review request with its selected harness, model, and workspace settings.
  2. NeMo Fabric validates the settings and starts the harness through its adapter.
  3. The harness performs the review using its own agent logic and tools.
  4. The adapter maps the outcome into a normalized result.
  5. NeMo Fabric returns the review output or failure details, plus available artifact and telemetry references.

For a single-request run, NeMo Fabric also attempts runtime cleanup before returning.

Use these guides to explore configuration, runtime ownership, and integration contracts: