Agentic ETL & Workflows

Avalanche is for workflows where an agent should contribute judgment, but should not silently own the whole path from input to action.

A hand-built application that calls an LLM directly gives you deterministic control, but makes every model contract, dependency, retry boundary, run, and output convention your responsibility. A hosted agent can absorb a broad task, but may make the execution path and critical actions harder to constrain.

Avalanche puts both in one workflow:

• Native agent steps handle interpretation, analysis, or flexible work.
• Deterministic Python nodes apply business logic and perform critical actions explicitly.
• The workflow DAG makes dependencies, parallelism, inputs, and outputs inspectable before a run starts.

The term agentic ETL describes this useful shape for data-oriented work. It does not require a fixed Extract/Transform/Load stage model or limit Avalanche to a particular storage backend.

​

Workflows make execution explicit

A workflow is a Python function decorated with @ava.workflow. Inside it, calls to decorated nodes build dependency edges rather than executing user code:

@ava.workflow
def review_feedback():
    return (
        (feedback := load_feedback())
        >> (
            (themes := extract_themes(feedback))
            & (risks := identify_risks(feedback))
        )
        >> (review := apply_business_rules(themes, risks))
        >> publish_report(review)
    )

>> expresses sequential edges and & groups parallel work. Capture a NodeFuture with := when a later branch or fan-in must reference it. At runtime, Avalanche schedules the graph topologically. A workflow can mix plain @ava.step nodes and @ava.agent_step nodes without treating agents as a separate control plane.

​

Start with a concrete run

Run Option A — Starter project (recommended), then use the customer-feedback example to see parallel agent work followed by deterministic reconciliation and artifact publication.