Skip to content

Latest commit

 

History

History
68 lines (51 loc) · 2.76 KB

File metadata and controls

68 lines (51 loc) · 2.76 KB

NVIDIA-labs Object Oriented Agents (NOOA) Adapters for NVIDIA NeMo Fabric

This directory provides two ways to run NOOA agents through the NeMo Fabric lifecycle contract. Choose the integration that matches the agent interface:

Integration Use It For
InteractiveAgent adapter Registered InteractiveAgent implementations, including CodingAgent and ArcSolverBase targets
BenchAgent adapter nooa_bench.BenchAgent tasks and Harbor evaluations

The shared nvidia.fabric.nooa adapter owns the common NOOA queue dispatcher. It also maps normalized skill paths and whole MCP servers into compatible registered targets. The dedicated nvidia.fabric.nooa.bench-agent adapter maps the benchmark-native BenchAgent task contract directly into a NeMo Fabric invocation.

Install

NOOA requires Python 3.12 or 3.13. Install NeMo Fabric, the adapters, and the tested NOOA packages together:

pip install "nemo-fabric[nooa]"

For a split adapter environment, install nemo-fabric-adapters-nooa[harness]. Use the adapter package without an extra when the environment already manages compatible NOOA packages. The installed wheel provides both adapter descriptors and the supported CodingAgent and ARC target descriptors; no source-path or explicit discovery configuration is required.

Configure Relay

Relay is optional. Install the adapter, tested NOOA packages, and compatible Relay Python package together:

pip install "nemo-fabric-adapters-nooa[full]"

Use nemo-fabric-adapters-nooa[relay] when the environment already manages the NOOA harness packages.

Both adapters declare Relay outputs for Agent Trajectory Interchange Format (ATIF), OpenTelemetry, and OpenInference. NeMo Fabric supplies the generated FABRIC_RELAY_CONFIG_PATH; the adapters reject ambient user or project plugin configuration and activate only the generated document.

A Relay setup failure prevents agent execution and returns a failed result. If artifact finalization fails after the agent completes, the adapter preserves the functional result and marks telemetry as degraded. A leaked Relay scope quarantines telemetry on later invocations.

Relay-backed Runtime.invoke_stream() returns raw Agent Trajectory Observability Format (ATOF) records and a separate terminal result. Neither adapter claims native model-response streaming.

Choose an Environment

These adapters do not provide a sandbox. Agents run model-generated code and shell commands with the permissions of their NeMo Fabric environment. Select an environment provider with the isolation required by your workload.