The fx adapter now runs the harness on the host

Built-in filesystem and shell operations retain their isolation within the sandbox, even with significant architectural changes to the fx adapter.

This adapter for the AI SDK's HarnessAgent now executes the harness and model requests directly within the same process as your application.

Leveraging libfx, the adapter streamlines operations; it no longer requires installing fx inside the sandbox, running a sandbox bridge, or exposing a sandbox port. Furthermore, model credentials now reside on the host, instead of needing forwarding into the sandbox environment.

The latest rewrite introduces several powerful capabilities for developers. These include built-in tool filtering, allowing you to select which native tools the agent can utilize.

It also supports mid-turn steering, enabling you to send fresh instructions while an existing turn is underway, and structured output generation matching a defined schema.

Moreover, the system can now return image tool results to the model via the toModelOutput() function.

Sessions possess the ability to be saved and subsequently resumed within a new host process.

To take advantage of this new adapter, developers must update to the latest version of @ai-sdk/harness-fx. Further details are available in the fx harness documentation.