Agent backends connect the compiler's generation task to a model or an external process. The driver selects the transport independently of the C++ source:

# Select a native API client, with credentials already configured.
$ llmc++ -fllm-backend=openai main.cpp -o main
 
# Select an installed, authenticated CLI adapter instead.
$ llmc++ -fllm-backend=codex main.cpp -o main
 
# Use the repository's custom-agent example.
$ llmc++ \
    -fllm-agent='python3 examples/python/18_agent.py' \
    -fllm-agent-config=examples/json/18_config.json \
    examples/18_custom_agent.cpp

The native clients run inside the C++ driver. CLI and custom agents use an external adapter or process. An explicit agent command takes precedence over backend selection.

All routes use the compiler's context tools and acceptance checks. Backend choice changes transport and model access, not whether a candidate must compile.