@sudoingX: The Harness Decides What the Model Sees

“The model does the thinking, the harness decides what it sees and what it remembers.”

@sudoingX spent the evening answering questions about xpress agent, the Rust harness he built for himself. Memory is the part he cares about most, he says: it keeps a capped memory and searches old sessions instead of stuffing them into context.

He likes pi’s minimal core, and says xpress agent goes the same lean way, “just in rust as one binary with no node_modules.”

Benchmarks are coming, he says, with the same model across the harnesses. For terminal speed he is capturing a pty with timestamps for the first frame, and checking flicker from the same captures.

On keys, his advice is to shrink what can go wrong. Keys that never leave your own machine can’t leak from someone else’s server, and a hard spend cap on every provider key limits the damage if one does.