Antonello Fratepietro

I wrote The Local Agent Trio: cmux + Pi + Unsloth Studio about splitting your local stack into cmux, inference, and Pi.

Ferrox is a pure-Rust inference engine for GGUF models.

Written August 2026, at v0.4.

In my previous article, The Local Agent Trio: cmux + Pi + Unsloth Studio, I split the local coding stack into three replaceable layers: cmux for the terminal workspace Unsloth Studio for inference Pi for the coding-agent loop That architecture still makes…

Running a local language model is easy right up until you want it to run well.

Cloud agents are convenient right up until you care about privacy, predictable costs, or a model that will not rate-limit you at 2am in the middle of a refactor.

Here is a failure mode I have hit more than once: an agent is halfway through a long task, then a deploy rolls the pods, a node gets evicted, or a human-in-the-loop step sits unanswered for six hours — and the whole run is just gone.

Every agent demo I have seen works beautifully.

Salvatore Sanfilippo — antirez, the person who gave us Redis — has a habit of building things that look obvious in retrospect and impossible before they exist.

I have spent more late nights than I would like to admit gluing inference stacks together, and the shape of the problem is always the same.