Amazon Web Services released Loom, an open source platform that deploys AI agents from fixed configurations rather than code generated at runtime, aiming to close a security gap enterprises face scaling agents.
The design choice sits at the center of what Loom is trying to solve. Many agent platforms generate code on the fly to handle each new request, but that code counts as untrusted and needs its own isolated environment to run safely, adding complexity for security teams. Loom instead builds agents once using AWS’s open source Strands Agents SDK and deploys them onto Amazon Bedrock AgentCore Runtime through configuration alone, so the underlying code can be reviewed and scanned a single time rather than re-checked with every deployment. AWS engineer Heeki Park, who introduced the platform, said it “gives builders the ability to easily create agentic applications” while keeping security and governance built into the deployment process itself.
On governance, Loom enforces three mandatory tags on every deployed resource, with room for organizations to add their own for cost tracking, and applies a two dimensional access model that pairs user roles with group tags to scope permissions. Credentials and behavioral rules are kept out of the platform entirely and pulled from AWS Secrets Manager only when needed, while inbound and outbound authentication runs through AgentCore Identity using OAuth2 and token exchange, so a user’s identity carries through an entire chain of agent requests rather than getting lost between steps. For actions considered sensitive, Loom adds human approval checkpoints built on the Strands Agents hook framework and Model Context Protocol elicitations, and it connects to the AWS Agent Registry, still in public preview, so agents and tools clear a governance review before reaching production.
AWS is offering Loom through AWS Labs on GitHub and inviting outside contributions, pitching it specifically at platform engineering teams building on fully managed AWS services rather than as a general purpose product. The release extends Bedrock AgentCore, which reached general availability in October 2025 after AWS first launched Bedrock Agents in late 2024, and arrives as enterprise agent tooling has become a crowded field, with OpenAI’s Agents SDK, Google’s Gemini agent framework and startups such as Cognition, maker of the Devin coding agent, all competing for the same enterprise deployments.
