What problem does it solve?
Integrating agent harnesses like Claude, Codex, Deep Agents, or Hermes Agent into an existing application, service, or evaluation platform requires translating your own configuration into NeMo Fabric's typed models, choosing the right runtime lifecycle, and handling normalized results correctly. This Skill guides that integration so you avoid unsupported patterns like importing internal modules or reimplementing harness control.
Core Features & Use Cases
- Config Translation: Map your application, job, or deployment config into an in-memory
FabricConfig using public models like HarnessConfig, ModelConfig, and ToolsConfig, with variants created via model_copy(deep=True).
- Lifecycle Selection: Choose between single-invocation
run(...), stateful multi-turn runtimes via start_runtime(...) with async with, native OpenAI streaming, or NeMo Relay ATOF streaming.
- Validation and Error Handling: Validate adapter selection and environment with
plan(...) and doctor(...) before execution, and consume normalized RunResult status, artifacts, telemetry, and FabricError subclasses.
- Use Case: A platform team adds NeMo Fabric to their evaluation runner so it can invoke Codex or Hermes Agent repeatedly with request correlation, artifact collection, and credential handling through environment variables.
Quick Start
Use the nemo-fabric-integrate skill to add NeMo Fabric to my Python service so it can run agent harness invocations through the typed SDK with validated configuration and normalized results.
