Source: AWS Open Source Blog — 2026-07-08

Summary

AWS open-sourced Loom, an enterprise-grade reference agent platform built on the Strands Agents SDK and Amazon Bedrock AgentCore Runtime. It provides a unified UI/API for agent lifecycle management, identity-provider integration, scope-based authorization (RBAC/ABAC), MCP and A2A integration management, discovery via the AWS Agent Registry, and three patterns of human-in-the-loop approval built on Strands hooks and MCP elicitations before an agent takes a sensitive action.

Key Takeaways

• Loom is a concrete, open-source ("paved path") reference architecture for enterprise agent governance, not just another single-purpose API.
• Built on Strands Agents SDK + Bedrock AgentCore Runtime, with identity-provider integration and scope-based RBAC/ABAC authorization baked in.
• Manages both MCP and A2A integrations centrally, plus agent discovery via the AWS Agent Registry.
• Ships three distinct human-in-the-loop approval patterns using Strands hooks and MCP elicitations, gating sensitive tool calls behind human sign-off.
• Code is public at github.com/awslabs/loom — a real, inspectable artifact rather than a whitepaper.

Reel Script

Hook Every enterprise wants AI agents in production, and every security team is terrified of what happens when one gets an unsupervised sensitive action wrong. AWS just open-sourced the blueprint they think solves that — and it's not another API, it's an entire governance layer.

Core Concept Agent governance isn't one feature, it's a stack: who is this agent allowed to act as, that's identity; what is it allowed to touch, that's scope-based authorization through RBAC and ABAC; which tools and other agents can it talk to, that's MCP and A2A integration management; and critically, does a human need to sign off before it does something risky. Think of Loom like the badge-and-checkpoint system in a secure building, except the employee is an AI agent — instead of one badge that opens every door, each agent gets scoped keys to only the doors it needs, with a guard standing at the doors marked sensitive.

Hands-On The concrete flow AWS ships: an agent authenticates through an identity provider, gets scoped RBAC/ABAC permissions, discovers other agents and MCP tools through the AWS Agent Registry, and before executing a flagged sensitive action, the request routes through one of three human-approval patterns built on Strands hooks and MCP elicitations, pausing the agent until a human approves or denies. It's all real, runnable code at github.com/awslabs/loom, built on Bedrock AgentCore Runtime — not a diagram in a slide deck, an actual reference implementation you can clone and adapt.

Takeaway My take: if your team is deploying agents into production without something like this identity-scope-approval chain, you don't have an agent strategy, you have a liability. Loom is worth cloning even if you don't run it as-is — steal the pattern. Follow for more on what production-grade agent infrastructure actually requires.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)