> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nemo/fabric/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo/fabric/_mcp/server.

# Release Notes

> Read the NVIDIA NeMo Fabric 0.1 release notes.

We're excited to announce NVIDIA NeMo Fabric 0.1, our first release. NeMo
Fabric introduces a typed, observable execution layer for running applications
across multiple agent harnesses. Applications, evaluation systems, and rollout
infrastructure can use the same configuration, lifecycle, result, artifact, and
telemetry contracts without building a separate integration for each harness.

## Release Highlights

### One Execution Contract Across Harnesses

NeMo Fabric provides one typed contract for configuring, launching, invoking,
and observing agent harnesses. Runs produce normalized responses, structured
failure details, lifecycle events, artifact references, and telemetry
references. This shared contract makes it easier to compare harnesses and build
experiment variations without rewriting each integration.

### Four Bundled Harness Integrations

The first release includes integrations for Claude Code, Codex, LangChain Deep
Agents, and Hermes Agent. Each adapter translates the shared NeMo Fabric
contract into the models, tools, skills, MCP servers, runtime state, and
observability capabilities supported by the harness.

### Persistent Multi-Turn Runtimes

Bundled adapters use persistent local runtime hosts that retain harness-native
session or thread state across turns. Single-invocation and multi-turn work use
the same lifecycle, with consistent startup, invocation, failure handling, and
cleanup behavior.

### Portable Configuration With Compatibility Validation

The typed configuration model covers models, instructions, runtime limits,
environments, tool policy, skills, MCP servers, telemetry, and
[NVIDIA NeMo Relay](https://github.com/NVIDIA/NeMo-Relay/tree/main).
Adapters declare the normalized configuration, telemetry, and runtime
capabilities they support. NeMo Fabric rejects incompatible normalized
settings and unsupported capability routes before harness invocation instead
of silently dropping them.

### NeMo Relay Observability and Live Streaming

All four bundled adapters integrate with NVIDIA NeMo Relay. Runs can produce
raw Agent Trajectory Observability Format (ATOF) events and normalized Agent
Trajectory Interchange Format (ATIF) trajectories, and export OpenTelemetry and
OpenInference through Relay. Relay-enabled runtimes can also stream live ATOF
records while keeping the normalized terminal result separate.

### Evaluation and Experimentation Workflows

The Harbor integration brings the shared NeMo Fabric contract into isolated
evaluation environments and returns normalized result and telemetry evidence.
This release includes a deterministic calculator workflow and a SWE-Bench
walkthrough built on a single Harbor task for comparing harness behavior on
the same task.

Guided notebooks and the experimentation CLI provide additional paths for
exploring harness and model variations, running maintained examples, and
creating editable Python or Rust projects.

### Cross-Platform Release Artifacts

NeMo Fabric ships Python artifacts for Linux on x86\_64 and arm64, macOS on
arm64, and Windows on x86\_64. The Rust core and experimentation CLI are also
published as independently versioned crates.

## Current Limitations

* Remote adapter services and a provider-backed third-party adapter registry
  are not included in this release.
* The configuration and adapter contracts use `v1alpha1` schema identifiers.