My favorite part about working on this was getting the chance to highlight AI Gateway’s best in class session affinity infrastructure. It handles things like tracking cache state for multiple models within a session, automatic fingerprinting, and much more. Cache hit rates can make or break inference economics, and AI Gateway helps you maximize them right out of the box. Any model, any provider.

I've been playing around with muse wonder via @vercel AI Gateway for agentic work and am super impressed with how well the model communicates. Here's an excerpt from me getting some help working on one of my motorcycles: - It starts with the lead clearly in an easy to read table and cites where it got the info from. - Next, it tells me exactly what I need to buy, and then gives me genuinely helpful bullet point tips without overdoing it (this is the taste part that's so tough to get right). - Finally, it suggests a follow up that actually makes sense. The tone and "feel" of this model just... clicks. This is all in fx.sh too, so the model is doing this by itself w/ minimal harness intervention. Haven't tested it for coding yet, but switching all of my @eve agents to use muse wonder as the orchestrator.

Incredibly bullish on this model, more to come!

i haven't written code locally in 2 mos. it's crazy what a good cloud agent setup (able to run projects locally, proper tool access, etc.) will do. and the orchestration layer is also all in the cloud! my mac is a chromebook :)

this guy ships.