AWS Introduces Loom — Open Reference Platform for Enterprise-Level AI Agent Management

Amazon Web Services (AWS) has introduced Loom, an open reference platform for building, deploying, and managing AI agents at the enterprise level. Loom uses the Strands Agents SDK to construct agents and the Amazon Bedrock AgentCore runtime for their execution. The platform includes a unified user interface and backend API, supports integration with identity providers, scope-based authorization, as well as agent lifecycle, memory resources, MCP servers, and agent-to-agent interaction (A2A). Loom is based on a prototype published by AWS Principal Solutions Architect Heeki Park in June and is now hosted in AWS Labs. The platform addresses seven key challenges of scaling agents, including enforcing resource tags, access control, propagating identity through delegation chains, and requiring manual confirmation for sensitive actions. Special attention is given to identity propagation: Loom implements a full authorization code flow for user interactions and uses RFC 8693 token exchange, supported by AgentCore Identity. Loom's deployment model avoids runtime code generation, instead using pre-written, configurable Python agents. Secrets and credentials are stored in AWS Secrets Manager. Management is handled through two mechanisms: mandatory resource tags and two-dimensional access control (role + group). The platform integrates with AWS Agent Registry (public preview), which complies with the A2A Agent Card and MCP Tool Architecture specifications. Manual confirmation is implemented through Strands Agents callback mechanisms. The community has responded to Loom with cautious optimism: on Reddit, no posts about actual deployments have appeared in two days. Loom is free and open source; costs are only associated with underlying managed services. It belongs to the rapidly emerging category of agent management platforms, unlike products such as Anthropic Claude Application Gateway, which focus on access control and cost for AI coding tools.