Skip to main content
Two agents take opposing positions on any topic you give them. A third agent judges the exchange, scoring each side on clarity, evidence, and persuasiveness. A good stress test for the injection guardrail — adversarial prompts tend to push at boundaries.

What it demonstrates

  • Per-agent guardrails overrides (injection on both debaters)
  • Sequential edges where each agent builds on accumulated context
  • Workflow-level and agent-level guardrail composition
  • The injection guardrail under combative, high-stakes prompt conditions

Run it

Workflow

docs/cookbook/adversarial-pair/workflow.yaml

Graph

Next steps

Guardrails

How injection detection and length limits work.

Conditional Pipeline

Route based on agent output using when: expressions.