About NeMo Fabric

Release Notes

View as Markdown

This page summarizes the current documentation-visible release state. GitHub Releases remain the source of truth for complete release history and tag-specific notes.

Current Release

NVIDIA NeMo Fabric 0.2 expands the execution contract from bundled agent harnesses to shared framework adapters and dedicated custom-agent adapters. Applications, evaluation systems, and rollout infrastructure can use the same configuration, lifecycle, result, artifact, and telemetry APIs across these Adapter Targets.

For release artifacts and the complete PR-by-PR history, refer to GitHub Releases.

Release Highlights

Public v1alpha2 Adapter Contract

The v1alpha2 contract makes the southbound adapter boundary independently consumable. This release publishes canonical JSON Schemas, dependency-free Python dataclasses with optional Pydantic interoperability, generated TypeScript types, maintained repository documentation, and a public adapter-authoring skill.

Adapters receive projected AgentConfig; northbound FabricConfig does not cross the supported adapter boundary. Each invocation receives typed AgentRunRequest and RuntimeContext values and returns one typed AgentRunResult; NeMo Fabric validates and enriches that result for consumers.

Harness and Custom-Agent Integration Patterns

The adapter contract now documents three integration shapes:

  • one shared adapter for many configurations of an opinionated harness;
  • one shared framework adapter for separately registered custom agents; and
  • one dedicated adapter for an application-owned custom agent.

The compact mini-SWE-agent adapter demonstrates the required lifecycle. The source-only NVIDIA NeMo Agent Toolkit reference demonstrates one shared adapter for calculator and email-phishing targets. The LangGraph email-phishing example demonstrates a dedicated custom-agent adapter with separate consumer, adapter, and agent code.

Target-Driven Discovery

Adapter Target Descriptors let a target package select its shared adapter and publish its adapter-scoped entry point and workflow settings schema. Consumers select a registered target with workflow.target_id; NeMo Fabric discovers metadata, validates compatibility, and projects the resolved entry point into AgentConfig before loading adapter code.

NeMo Fabric discovers descriptors bundled with the runtime, installed below the shared package-data root, or listed explicitly in FabricConfig.discovery.local_paths. Conflicting records with the same ID fail planning instead of silently overriding one another.

Expanded Configuration and Streaming

The normalized contract adds named tool and tool-group definitions, per-server MCP tool filters, and MCP OAuth 2.0 and service-account authentication metadata. Adapters advertise only the normalized fields and authentication modes they apply, and unsupported configured behavior fails planning.

Native OpenAI Chat Completions streaming is available to adapters that declare and implement the optional capability. Relay-backed Agent Trajectory Observability Format (ATOF) streaming remains the primary normalized streaming API and does not require another adapter method. This release also updates NVIDIA NeMo Relay integration to 0.7.

Bundled Adapter Alignment

Claude Code, Codex, Hermes Agent, LangChain Deep Agents, and mini-SWE-agent use the v1alpha2 descriptor, configuration, request, context, and result boundary. Each adapter publishes its supported configuration, settings schemas, runtime requirements, telemetry outputs, and optional capabilities for planning and diagnostics.

Current Limitations

The current release has the following limitations:

  • Provider-backed third-party adapter registries and remotely hosted adapter services are not included.
  • The automated adapter conformance suite is not included. Adapter authors can use the maintained verification checklist for the required profile and declared capabilities.
  • Cancellation, live updates, and service lifecycle flags are reserved until a runtime binding exposes and tests the corresponding adapter operations.