Introducing Run SDK: Safe eval for Agents

Running code written by AI agents for tool orchestration with the existing eval poses security risks by exposing application secrets and access to internal services. To address this, Vercel has launched the Run SDK.

The Run SDK executes untrusted JavaScript and TypeScript within a QuickJS context inside worker threads, blocking direct access to Node.js or the network. Applications expose only the necessary capabilities to the sandbox via hostFunctions, with calls handled through serialization.

This SDK is also utilized as an internal module supporting code-mode tool execution in the AI SDK. Agents can handle multiple service calls and logic simultaneously within a single model response, enabling efficient task completion without injecting unnecessary data into the context.

Additionally, it provides functionality to pause execution and wait for a decision before resuming for sensitive tasks that require authentication or human-in-the-loop approval.