The fx adapter now runs the harness on the host
Vercel’s AI SDK fx adapter now executes the harness and model requests in the same process as the host application, improving efficiency while maintaining sandbox isolation for filesystem and shell operations.
The fx adapter for the AI SDK’s HarnessAgent has been updated to run the harness and model requests within the host application’s process, eliminating the need for separate sandbox execution. This change leverages libfx, a TypeScript library for fx, to streamline operations. The adapter no longer requires installing fx inside the sandbox, removing the dependency on a sandbox bridge or exposed port.
Model credentials are now retained on the host system instead of being forwarded into the sandbox, enhancing security by keeping sensitive data outside the isolated environment. Sessions can be saved and resumed in a new host process, improving flexibility for developers managing long-running tasks.
The update removes the requirement for a sandbox bridge, simplifying the architecture and reducing overhead for developers using the AI SDK. Built-in filesystem and shell operations remain isolated in the sandbox, ensuring security without compromising functionality.
To adopt the new adapter, users must update to the latest version of @ai-sdk/harness-fx. The change is designed to improve performance and developer experience while maintaining the security model of the AI SDK’s HarnessAgent.