Avalanche home pagelight logo
  1. Build Workflows
  2. Declare agent steps

@ava.agent_step is an ordinary workflow step with an injected callable agent. The body maps workflow values into model inputs, awaits a raw prediction, and returns an explicit result.

​
Use a typed signature for shared contracts

import avalanche as ava
from pydantic import BaseModel


class Review(BaseModel):
    summary: str
    approved: bool


class ReviewDocument(ava.Signature):
    """Review a document for publication."""

    document: str = ava.InputField(desc="Document text to review.")
    review: Review = ava.OutputField(desc="Publication decision and summary.")


@ava.agent_step(ReviewDocument, lm="openai/gpt-5.6-terra")
async def review_document(document: str, *, agent: ava.Agent) -> Review:
    prediction = await agent(document=document)
    return prediction.review

The agent parameter is required and keyword-only. Workflow callsites supply only normal workflow values; Avalanche injects the agent when the node runs.

ava.Signature, ava.InputField, and ava.OutputField align with DSPy signature conventions. Native DSPy signatures are accepted too.

​
Use an inline signature for a local contract

answer_signature = ava.agent.Signature(
    "question: str, context: str -> answer: str",
    "Answer the question from the supplied context.",
)


@ava.agent.step(answer_signature)
async def answer_question(question: str, context: str, *, agent: ava.Agent) -> str:
    return (await agent(question=question, context=context)).answer

Use a typed class for a substantial or reusable contract. Use the inline form for a compact, local one.

​
Add skills and tools at the step

Skills and tools are capabilities of a particular agent step, not properties of the signature:

@ava.agent_step(
    answer_signature,
    skills=[ava.agent.skills.pdf],
    tools=[search_contract_repository],
)
async def answer_contract_question(question: str, *, agent: ava.Agent) -> str:
    return (await agent(question=question)).answer

Tools are ordinary callables with unique stable __name__ values. Avalanche ships lazy PredictRLM skill re-exports such as pdf, docx, and spreadsheet.

​
Set runtime defaults deliberately

Shared model settings belong on the workflow; a setting on a specific agent step takes precedence:

@ava.dest
def publish_review(review: Review) -> Review:
    return review


@ava.workflow(
    agent_defaults={
        "lm": "openai/gpt-5.6-terra",
        "sub_lm": "gemini/gemini-3.5-flash",
        "max_iterations": 30,
    }
)
def proposal_flow():
    return (review := review_document(ava.input.document)) >> publish_review()

signature, skills, and tools remain step-local. Agent steps are quiet by default; set verbose=True only when live PredictRLM trace output is needed.

Next: provide runtime input and artifacts.