AWS launches InvokeGuardrailChecks API for agentic AI safety
Apply per-step safety checks without creating guardrail resources in agent loops.
Amazon Bedrock Guardrails has announced the InvokeGuardrailChecks API, a new capability designed to inject targeted safety checks into agentic AI applications without the overhead of managing separate guardrail resources. Unlike traditional guardrail patterns that require a resource creation step, this API works in a resourceless, detect-only mode, returning numeric scores for each safeguard applied. Developers can then define custom thresholds and actions—such as block, bypass, retry, or log—directly in their application logic. This is crucial for multi-turn AI agents that iterate loops of user input, tool calls, and model responses, each with a different risk profile.
The API uses a structured messages schema with roles (system, user, assistant) to evaluate content in context. For example, a customer support agent might need to check for prompt injection early, sensitive PII mid-turn, and harmful content at output. With InvokeGuardrailChecks, each turn can run only the relevant checks, and because no guardrail IDs or versions need to be managed, scaling across hundreds of agents becomes practical. This approach reduces operational complexity while maintaining granular safety controls, making it a significant step for production-grade agentic AI deployments.
- The InvokeGuardrailChecks API is resourceless: no CreateGuardrail step, no version management, and each request specifies which safeguards to run.
- It returns numeric scores for each safeguard, enabling custom thresholds and actions (block, bypass, retry, log) in application logic.
- Uses a structured messages schema with roles (system, user, assistant) to accurately evaluate content in multi-turn agent loops.
Why It Matters
Granular, per-turn safety for AI agents without resource overhead—critical for scaling production agentic workflows.