Examples and References

Use the example that matches the integration shape you chose. Review its descriptor, lifecycle implementation, and consumer configuration, in that order. The canonical schemas remain authoritative when an example omits a field.

Harness Adapter

Use the Hermes Agent adapter as the primary reference for integrating an opinionated harness. Its descriptor and runtime demonstrate normalized models, MCP, instructions, runtime settings, persistent session state, and Relay telemetry through one harness adapter.

Use mini-SWE-agent as the secondary minimum-surface reference when you want to understand the required boundary at a glance. Its package contains:

• A static Adapter Descriptor that accepts a small normalized config surface
• One runtime class with start, invoke, and stop
• The optional common Python lifecycle host
• One package README with a complete consumer configuration

The adapter retains the model, environment, agent, and conversation state in one runtime instance. It does not implement MCP, skills, telemetry, or native streaming, which keeps the required path visible.

Shared Adapter With Registered Custom Agents

Start with the NeMo Agent Toolkit reference when one adapter must support many custom agents built for a framework. Read these files in order:

1. nat.fabric-adapter.json declares target_types: ["workflow"], normalized configuration, tool definition schema, and native OpenAI streaming.
2. email-phishing-analyzer.fabric-target.json selects the shared adapter and supplies the workflow entry point and settings schema.
3. calculator.fabric-target.json shows a second target that reuses the same adapter contract.
4. adapter.py translates AgentConfig into NeMo Agent Toolkit configuration and owns the retained workflow lifecycle.

The consumer selects a target, not an entry point:

1
FabricConfig(
2
    workflow=WorkflowConfig(
3
        target_id="nvidia.examples.nat.email-phishing-analyzer",
4
        settings={"llm_name": "default"},
5
    ),
6
    models={"default": ModelConfig(...)},
7
)

Planning resolves the target’s adapter and projects the target-owned entry point into AgentConfig.workflow. A second target changes the target ID and settings schema without changing the shared Adapter Descriptor.

Dedicated Custom-Agent Adapter

Start with the LangGraph email-phishing example when the custom application is the clearest integration boundary. Its directories make ownership explicit:

• consumer/ owns FabricConfig and invocation.
• adapter/ receives AgentConfig, translates dependencies, and owns the lifecycle.
• agent/ contains application behavior without an NVIDIA NeMo Fabric dependency.

The descriptor directly selects this adapter and does not declare target_types or accept workflow. Its minimum path maps one model and one system instruction, compiles the graph during start, invokes the retained graph, and releases its references during stop. Optional MCP and Relay modules are adjacent additions rather than requirements of the minimum lifecycle.

Follow the Type Boundary

Across all examples, keep these ownership boundaries intact:

Use the schemas/adapter-contract/ directory for exact southbound wire shapes and Stage 1 to begin your own descriptor.