Lanes

A lot of what you send a coding agent is easy - “yes, push it”, “commit it”. Sending that to a cheap model sounds smart, but every switch kills the prompt cache. Naive switching cost us 40% more. So we built Lanes, a Pi extension. SubQ Snap sorts each message into a lane with its own Meta Muse model. The big model never switches - cheap turns get appended to its thread, so its cache stays intact. Easy turns got 65% cheaper at about the same quality (11.5/12 vs 12/12). On two hard tasks, our first version scored 2/5 and 0/7; after the fix, 5/5 and 7/7. It also gets better: repeated git requests go to a 1.5B model learned from requests mined from our own sessions. It scored 100% on held-out tests (its cloud teacher: 65%), and a weaker version was rolled back. https://claude.ai/artifact/MsLRftvtL428kM6kMcLd82