Skip to content

Guardrails

ThinkWork uses Amazon Bedrock Guardrails as the primary enforcement layer for managed agents.

Guardrails are evaluated around model invocation and can:

  • block disallowed topics
  • filter unsafe content
  • detect or redact PII
  • apply grounding checks
  • enforce word or phrase filters

Guardrails are referenced from agent templates, which lets you apply the same safety boundary across a fleet of agents.

That keeps enforcement in one place instead of scattering it across prompts.

When a guardrail blocks or alters a response, the resulting event is visible in ThinkWork’s thread and control surfaces.

This matters for debugging because you can distinguish:

  • model behavior
  • tool behavior
  • policy enforcement