guardrails-block.yaml

apiVersion: agent-controller.dev/v1alpha1
kind: Agent
metadata:
  name: guardrails-block-demo
  owner: ccdevelopforfun
  description: |
    Demonstrates the default hallucination-detector mode. When the model
    fabricates tool-call XML (<invoke>, <function_calls>, etc.) in its
    message text, the runtime emits a wire `error` event and ends the
    session with reason=error and a non-zero CLI exit code.

    To provoke the failure deliberately, this spec assigns a task that
    typically requires a tool the agent does NOT have. The model is most
    likely to refuse honestly thanks to the v0.1.8 honesty preamble; if
    it instead writes fake <invoke> XML, the block guardrail catches it.

spec:
  model:
    provider: anthropic
    name: claude-sonnet-4-20250514
    temperature: 0.2
  persona:
    role: Demo agent without any tools
    instructions: |
      You have no tools. Always be honest about what you cannot do.
  task: |
    Look up the current weather in Tokyo and return the result as JSON.
  tools: []
  runtime:
    type: local
  guardrails:
    hallucinationDetector: block