Vercel rewrites its fx coding-agent harness to run on the host
The fx adapter for the AI SDK now keeps model credentials on the host instead of forwarding them into the sandbox, and gains tool filtering and mid-turn steering.
Vercel rewrote the fx adapter for HarnessAgent, the part of its AI SDK that wraps coding-agent harnesses, so the harness and model requests now run on the host process rather than inside the sandbox. Only filesystem and shell operations stay isolated in the sandbox, and model credentials stay on the host instead of being forwarded in.
The change, published 27 October, means fx no longer needs to install itself inside the sandbox, run a sandbox bridge, or expose a sandbox port. Three capabilities arrive with the rewrite: built-in tool filtering, so you choose which tools an agent can use, mid-turn steering, to send new instructions while a turn is running, and structured output, for responses that match a schema. Sessions can also be saved and resumed in a new host process. Anyone using the fx harness needs to update @ai-sdk/harness-fx to the latest version to pick up the new adapter.
Why it matters: if you use the fx harness, credentials no longer have to leave the host process at all, narrowing what a compromised sandbox can reach, and setup gets simpler without a sandbox bridge or exposed port to manage.