July 21, 2026 — Amazon Web Services has released Loom for AWS, an open-source, enterprise-grade platform that tackles a problem now keeping platform engineers up at night: how to deploy AI agents at scale without inviting chaos, security holes, or a compliance nightmare. And in classic AWS fashion, the answer is not a managed service but an opinionated reference implementation—available on AWS Labs—that enterprises can fork, customize, or simply learn from.
If you’ve been following the agentic AI wave, the pitch sounds familiar. Build, deploy, govern. What sets Loom apart is how deeply it embeds governance into every step, treating it not as an afterthought but as the foundational layer. Three mandatory tags (loom:application, loom:group, loom:owner) are stamped on every resource. Access control blends role-based and attribute-based models—RBAC and ABAC—so permissions can be sliced by group or function. And the platform flat-out refuses to let agents generate or execute code at runtime. Instead, a pre-written, configurable Python agent built with the Strands Agents SDK gets deployed repeatedly; only its configuration changes. “Code never changes between deployments,” the AWS team explains, meaning security teams can scan it once and move on.
The architecture grew out of an internal prototype by Heeki Park, a principal solutions architect at AWS, before being promoted to AWS Labs in June. Since then, the project has seen active commits, with a fresh deployment clocking in at roughly 40 minutes on a clean AWS account—a number that impressed early explorers on Hacker News. “Forty minutes to a fully governed, stateful agent system. That’s insane compared to stitching together LangChain, Bedrock, and a dozen IAM policies by hand,” one commenter wrote. Another added a note of caution: “But only if you’re all-in on AWS. If you ever plan to leave, you’re re-architecting from scratch.”
That tension between governance convenience and platform lock-in runs through every conversation about Loom. The platform leans heavily on Bedrock for model inference, AgentCore for agent infrastructure, and Secrets Manager for credential injection. When an agent acts on behalf of a user and hops through MCP servers and REST APIs, Loom implements the full OAuth2 authorization code flow with RFC 8693 token exchange to preserve the original user’s identity. That’s an elegant solution to the delegated-identity problem, but it also ties the platform tightly to AWS’s IAM and AgentCore Identity. For enterprises that have already standardized on AWS, that’s a feature. For those hedging between clouds, it’s a serious dependency.
Developers poking around the GitHub repository have surfaced early quirks. One issue documents a builder agent getting stuck in an Opus 4.6 configuration wizard. Another asks for better LLM resource usage tracking, something the platform claims to support but that early testers found lacking. Still, the mandatory tagging feature alone has drawn praise from compliance circles. “Being able to answer ‘who owns this agent and what’s it costing us?’ without a 10-page report is rare in this space,” a cloud architect shared on a DevOps Slack. “Loom enforces it automatically. That’s the kind of opinionated design that actually works in enterprises.”
[SPONSORED]
▶ ENTERPRISE GPU CLUSTERS ◀
Scale your AI model training seamlessly. Book a Demo.
Speaking of costs: because Loom is open source, there’s no platform license fee. You pay only for the underlying AWS services—Bedrock inference, AgentCore runtime, storage, and data transfer. That can be cheaper than LangGraph Cloud’s per-action pricing at scale, but it also means you’re responsible for running and maintaining the control plane. The break-even point compared to a fully managed SaaS probably sits somewhere between 15 and 18 months, similar to other self-hosted versus cloud debates. The real savings, though, might come from avoiding the governance overhead that cause so many agent pilots to stall. IDC recently warned that “the scale of agent deployments must be accompanied by AI-ready data foundations, a multi-agent orchestration platform, and a sound governance mechanism”—words that could have been cribbed from Loom’s design brief.
Not everyone is convinced that AWS’s open-source play is entirely altruistic. By releasing Loom as a reference platform, the company sets the standard for what agent governance should look like on AWS, and quietly nudges teams toward Bedrock and AgentCore. “It’s the same playbook as OpenSearch,” noted a cloud analyst on Twitter. “Open source the blueprint, make consumption of your services the path of least resistance, and call it customer choice.” AWS, for its part, insists Loom is meant as a concrete example, not a product. Organizations can take the code and build their own platforms, with or without AWS.
Whether Loom becomes a community-supported staple or a niche project for AWS-centric shops remains to be seen. The platform already supports MCP servers, A2A integrations, and LiteLLM proxies for alternate model providers—so you’re not locked into a single LLM. But the infrastructure layer is distinctly Amazon. As one GitHub comment put it: “It’s a beautiful cage, but a cage nonetheless. If you’re already living in the AWS zoo, you’ll love it.”