Michael Mignano

Three catalysts I have in mind for the growth of callable models. Each increases the need for intelligent routing, but perhaps more importantly (and valuably) open harnesses : 1) personalized agents that hold private information (what's better to call for insight on XYZ, opus 4.7 or opus 4.7 wrapped in context with private info from the worlds foremost expert on XYZ) 2) domain-specific models like Cursor's Composer 2 who's value prop isn't beating performance but being on-par with performance at a cheaper cost. as token spends become a larger and larger portion of expenses for businesses, the value from routing to smaller, dedicated models with on-par performance will grow with it 3) domain-specific models who's value prop is being more performant. we see these in domains with sufficiently idiosyncratic logical structures / input modalities (bio, drug discovery, physics,...) In order to stay at the frontier, companies will want tools that bridge context across these expert models But I don't think value accrues to routers. The most successful router of all time, Google, which routes users to their desired webpages, is notably not just a router. It's also a "harness" in some sense, that the user directly interacts with. The router is just the algorithm that's called after the user's interaction, in some sense it's just an API that serves as middleware. The harness logs info about the users preferences (time staying on the site, what link they clicked, whether they returned for another similar search...). The harness then coordinates any updates to the router's config next time around. As time went on, Google relied less and less on the core routing algorithm, PageRank, and more and more on ML algos trained on the user feedback logged. The harness, which is the point of contact for the user, controls the user feedback, and so it controls the power to verticalize into routing (and maybe even further into the model). This was how cursor played out. It started as a code-specific harness (agentic IDE), it introduced routing (auto), and eventually trained it's own model (composer 2). This evolution was uniquely enabled by it's positioning in the line of user feedback.