Vercel has changed where its fx coding agent keeps its brain. In a changelog post by Felix Arntz, Vercel says the HarnessAgent adapter for fx, in its AI SDK, is rebuilt on libfx, a TypeScript library, so the harness (the software wrapped around the model) and the model requests run inside your application’s own process. The agent’s file and shell commands still execute in the isolated sandbox.
Setup gets lighter. Teams no longer install fx in the sandbox, run a bridge into it, or open a sandbox port. Model credentials sit on the host, so the code the agent runs never sits next to the keys, which is the sensible place to draw that line.
The rewrite also adds tool filtering, so you pick the built-in tools the agent may call. You can steer it with fresh instructions mid-turn, get answers in a fixed schema, and hand images back to the model via toModelOutput(). A saved session can pick up again in a different host process. The post gives no performance figures.
Anyone already on @ai-sdk/harness-fx gets all of this by updating, and any sandbox bridge they built can come out.
Reported by the Vercel changelog (post by Felix Arntz), which carries no publication date in our copy.
