Skip to main content
Two agents connected by an edge. The first classifies the user’s intent; the second reads that classification from the workflow context and composes a reply. This classify → respond pattern appears in nearly every customer-facing agent workflow.

What it demonstrates

  • Connecting two nodes with an edge
  • Writing intermediate results to working.* context paths
  • Mixing providers: OpenAI classifies, Anthropic replies
  • Workflow-level guardrails

Run it

Workflow

docs/cookbook/sequential-pipeline/workflow.yaml
This workflow mixes providers: OpenAI handles classification, Anthropic writes the reply. Each agent independently resolves its own provider at runtime.

How data flows

  1. classify receives the user’s message and writes a single intent label (e.g. "complaint") to working.intent.
  2. reply receives the output of classify as its user message and composes the final response.

Graph

Next steps

Conditional Pipeline

Route to different handlers based on the classification.

Telephone Game

Chain five agents and watch meaning drift.