Pipelines

Pipelines combine deterministic TypeScript, agents, and extractors into a typed, inspectable workflow.

Explore pipelines

Page	Learn how to
Typed input	Validate input before any work begins.
Steps	Transform values with synchronous or asynchronous logic.
Agents and extractors	Add model reasoning and schema-validated extraction.
Composition	Reuse pipelines and name inspectable stages.
Parallel and batch	Run independent branches or many inputs safely.
Runs and errors	Execute, observe, inspect, and map failures.
Production workers	Move long-running workflows out of HTTP requests.

The pipeline flow

Validate input → run typed stages → return final output

A pipeline is useful when work is more than one prompt: normalize input, load product data, ask an agent, extract fields, run independent checks, and shape the final result.

Use ordinary TypeScript first

Keep normalization, authorization, database access, branching, side effects, and final response shaping in .step(...). Use .prompt(...) only for model reasoning and .extract(...) only when existing content must become structured data.

If one function already expresses the workflow clearly, it does not need a pipeline. Reach for a pipeline when typed composition, reusable stages, bounded batch work, or graph inspection adds practical value.