2. Switch Harnesses
This second quickstart demonstrates the central reason to use NVIDIA NeMo Fabric: your application can keep its execution and result-handling contract while changing harnesses. Complete 1. Install One Harness first and retain its config in first_agent.py.
Install the Second Harness
This example compares Deep Agents with Hermes Agent, using its ready-made NeMo Fabric adapter. The supported Hermes v0.22.7 stack requires Python 3.14; NeMo Fabric itself supports Python 3.11–3.14. Keep Hermes in a separate environment to isolate its dependencies. Both variants use the NVIDIA model and credential from the first quickstart.
With the first quickstart’s .venv still active, install uv and Git, then run the following in your application directory using a POSIX shell. The commands check out the supported Hermes release, use its frozen dependency lockfile, and install the Hermes adapter version matching your application:
Do not activate .venv-hermes; your application still runs in .venv. HERMES_ADAPTER_PYTHON selects the second interpreter only when launching Hermes. If you already have the supported Hermes environment, install the matching adapter there and set this variable to its absolute interpreter path instead. Refer to the Hermes installation guide for the source-checkout alternative and upstream platform guidance.
Construct Explicit Variants
Save the following as compare_agents.py. Set HERMES_ADAPTER_PYTHON to the supported Hermes environment’s absolute Python path before running it:
The environment-variable switch is sequential and intended for this local lesson. Do not mutate shared process environment variables around concurrent runtime starts. Production launchers should establish consistent execution environments.
Interpret the Results
Look for two labeled results, one per harness. Each should report Status: succeeded, the expected adapter ID, and a review identifying that an empty input list causes division by zero. For example, the output can look like this; review wording varies:
The shared RunResult fields let your application check status before reading output.response, regardless of the harness. If a result fails, inspect the labeled error and follow Troubleshooting; configuration or startup failures can instead raise SDK exceptions. Different prompts, defaults, tools, and native semantics can produce different reviews. This exercise demonstrates common invocation and result handling; it is not a benchmark or proof of equivalent behavior.
Before changing another configuration field, consult Harness Compatibility. For a controlled task and verifier, use Harbor. The maintained variations notebook expands this pattern to additional harnesses, capabilities, and telemetry.
Keep HERMES_ADAPTER_PYTHON set and continue to 3. Add Skills and MCP to extend the Hermes variant using the same model and code-review task.