Configure models and providers

Avalanche sends agent-model requests through LiteLLM. Configure provider credentials in the environment of the process running the workflow, then use a provider-qualified model identifier on an agent step or at the workflow level.

​

Set a model on one step

@ava.agent_step(
    ExtractThemes,
    lm="openai/gpt-5.6-terra",
    sub_lm="gemini/gemini-3.5-flash",
)
async def extract_themes(feedback: str, *, agent: ava.Agent) -> ThemeReport:
    return (await agent(feedback=feedback)).report

lm selects the main model and sub_lm selects the sub-model.

​

Share defaults across a workflow

@ava.workflow(
    agent_defaults={
        "lm": "openai/gpt-5.6-terra",
        "sub_lm": "gemini/gemini-3.5-flash",
    }
)
def feedback_workflow():
    return load_feedback() >> extract_themes()

A value on an individual agent step overrides the workflow default. Workflow agent_defaults configure runtime policy; signatures, skills, and tools remain explicit on each agent step.

Resolution order:

agent-step runtime kwargs
  → workflow agent_defaults
  → Avalanche agent defaults
  → PredictRLM defaults

​

Keep credentials out of workflow code

Use environment variables recognized by your LiteLLM provider. In a starter workspace, scripts/configure-provider.sh writes needed credentials to the Git-ignored local .env file. Do not put provider secrets in workflow modules, source control, or execution-service requests.

See declare agent steps for signatures, skills, tools, and model contracts.