AI Agent Governance on AWS: Block Agents, Prove EU AI Act Compliance

This article explores the practical challenges of governing multi-agent AI systems on AWS, specifically focusing on the requirements of the EU AI Act. The author demonstrates how to implement runtime guardrails and audit evidence using Amazon Bedrock and the Traccia platform. By building a synthetic loan-decision crew, the guide illustrates how to perform prompt injection blocking, PII redaction, and compliance stamping. A key takeaway is the distinction between local SDK-based governance and platform-level enforcement. The author highlights that effective governance requires scoping policies to the specific sub-agents performing actions, rather than just the entry point. This technical deep dive provides a reproducible framework for developers looking to satisfy regulatory obligations like Article 12 and 14 of the EU AI Act without needing to overhaul their entire application architecture, while offering an honest critique of where automated governance tools succeed and where they require manual configuration.