What sandboxing an AI coding agent in a VM costs
- Source
- veloworkspaces.com
- Author
- michael_luog
- Date
Context
A coding can run its commands inside a virtual machine while the AI model stays on the host computer. A published experiment on an M4 Mac mini connects the two with a small bridge: a proxy in the virtual machine forwards requests over a virtual socket to a listener on the Mac, which passes them to the model engine still running there. With eight simultaneous clients, total time in the virtual machine was about 5% longer than running directly on the host with MLX, but about 15% shorter with Ollama. The two engines were measured in separate sessions, and the author's explanation for the Ollama result is a working theory. The measurements show that the cost of this arrangement varied with the serving engine on this hardware, which is a reason to measure the intended setup. They do not establish a general performance penalty or benefit, or any security guarantee.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
Checking sign-in…
Loading comments…