Manvitha Potluri, CITP, is a DevOps Cloud Solutions Architect at 24X7 Systems Inc. with 11+ years of experience in cloud infrastructure, DevOps, and AI systems engineering. Manvitha Potluri's work focuses on AWS, EKS/Kubernetes, Terraform, and secure agentic AI, including a patent-pending multi-layer guardrail architecture for AWS Bedrock and open-source infrastructure modules published on the HashiCorp Registry. Manvitha Potluri contributes to CNCF ecosystem projects, publishes technical books, and holds a Master of Science in Computer Science Engineering.

A Five-Layer Guardrail Architecture for Secure Agentic AI Systems on AWS Bedrock

Abstract : As enterprises move from conversational LLMs to autonomous agents that invoke tools and act on production systems, single-layer content filters are no longer sufficient. This talk presents a five-layer guardrail architecture for agentic AI workloads on AWS Bedrock, combining input validation, policy-based content filtering, contextual grounding checks, output verification, and pre-execution authorization of agent actions.
The architecture was empirically evaluated against baseline single-layer Bedrock guardrails using benign, borderline, and adversarial prompt sets. The five-layer configuration achieved 94% overall accuracy, a 6% false-positive rate, a 0% false-negative rate, and 100% detection of adversarial prompts, while reducing mean latency 7.6× (1,828 ms vs. 13,874 ms) through layered short-circuit evaluation.
The talk covers the design rationale for each layer, lessons from evaluation (including how API throttling can silently corrupt benchmark results), and how the architecture is packaged as a reusable, open-source Terraform module on the HashiCorp Registry for repeatable deployment. It closes with an extension currently in development: an "agentic action gate" that enforces authorization and budget limits on agent tool calls before execution.
Attendees will leave with a practical blueprint for deploying governed, auditable agentic AI in regulated cloud environments.
Keywords: Agentic AI, LLM guardrails, AWS Bedrock, AI governance, DevSecOps, Terraform