Vercel's fx adapter now runs the AI SDK harness on the host, not in the sandbox

Harness and model requests move to the host

Under the new design, the harness and the model requests share a process with your application. The built-in filesystem and shell operations remain isolated in the sandbox, so the split is between orchestration on the host and tool execution in the sandbox.

The adapter is built on libfx, the TypeScript library for fx. According to Vercel, this removes several setup steps that the previous approach required. The adapter no longer needs to install fx inside the sandbox, run a sandbox bridge, or expose a sandbox port.

Model credentials also change location. They now stay on the host rather than being forwarded into the sandbox. For teams that run agent-generated commands in an isolated environment, this means the sandbox no longer has to hold the keys used to call the model.

New capabilities in the rewrite

Vercel says the rewrite adds several features alongside the architectural change:

• Built-in tool filtering: developers can choose which built-in tools the agent is allowed to use.
• Mid-turn steering: new instructions can be sent while a turn is still running.
• Structured output: the agent can generate responses that match a schema.
• Image tool results: images can be returned to the model through toModelOutput().

Sessions can also be saved and resumed in a new host process. Because the harness now lives in the application process, a session is no longer tied to the lifetime of a single host process.

How to get it and what it means

To use the new adapter, developers need to update @ai-sdk/harness-fx to the latest version. Vercel points to the fx harness documentation for further details.

The changes mainly affect developers who build agents with the AI SDK's HarnessAgent and run them with fx. Fewer moving parts in the sandbox, with no bridge and no exposed port, simplifies the setup described in the announcement. Keeping model credentials on the host narrows what the sandbox has access to. The added controls, such as tool filtering and mid-turn steering, give developers more say over what the agent can do and how it can be directed while it works.

The source does not give performance figures or details on how existing setups should migrate beyond updating the package, so teams with current deployments should consult the documentation.