AWS Open-Sources Loom, a Reference Platform for Securing AI Agents at Scale

AWS has released Loom, an open-source reference platform that provides a blueprint for how enterprises can manage AI agents across their infrastructure with security controls embedded throughout. The platform combines a management dashboard and API layer with authentication integration, permission scoping, and lifecycle controls for agents, memory systems, protocol servers, and agent-to-agent connections. Rather than a finished product, AWS positions Loom as a concrete reference implementation that engineering teams can study and adapt to build their own agent governance layers.

The Core Governance Challenge#

Building and deploying AI agents at enterprise scale creates several interconnected governance problems: maintaining consistent naming conventions across resources, implementing granular permission controls, standardizing deployment processes, ensuring code quality before production, tracking which user is ultimately responsible when agents delegate tasks through chains of services, controlling agent proliferation, and enforcing human approval gates for high-risk operations.

The most technically demanding of these involves identity propagation. When an agent performs work on a user's behalf and passes requests to intermediary services (such as MCP servers) that themselves call downstream APIs, each layer in the chain needs an access token that correctly reflects the original user's identity and their authorized permissions. Loom solves this by implementing the complete OAuth authorization flow for initial user interactions and leveraging the RFC 8693 token exchange standard. This ensures that as requests move through the agent to intermediary services to final API endpoints, the originating user's identity remains visible in every token, allowing downstream systems to enforce proper access controls.

Configuration-First Deployment#

Loom deliberately avoids code generation at runtime. Instead, it packages a pre-built, configurable Python agent that can be deployed repeatedly with different behavior rules, memory configurations, and service integrations supplied at deployment time. The agent code itself never changes—only the configuration does. This allows platform teams to review the code once in detail, add enterprise-specific customizations like audit logging, and then reuse that validated agent across all deployments without re-reviewing.

For teams without customization needs, Loom also provides a no-code deployment path through pre-built tooling. Notably, secrets and credentials are never stored within Loom itself; instead, they live in AWS Secrets Manager and are fetched on-demand, with authentication handled by the underlying identity layer.