Discussion about this post

User's avatar
Foundation's Edge's avatar

Useful survey — the spread from learned routers to gateway rules to Devin's sidekick pattern shows routing dissolving into ordinary platform engineering. The piece those systems still need is an acceptance gate: a route is a promise about who decides the output counts, and most portfolios leave that implicit. Treating routing as infrastructure, with seats, prices, and the two ledgers of capability and cost: https://seldondance.substack.com/p/production-ready-llm-routing

Foundation's Edge's avatar

Calling this an intelligence control plane rather than a router is the right frame — and the survey shows the winners still disagree on what that plane should optimize: OpenRouter Fusion spends extra inference to cut uncertainty, Vercel encodes deployment policy, Splunk chases cost per step. That disagreement is the tell that routing is a systems design problem, not a config default. Our production routing breakdown (multi-signal, learned, semantic, managed) reaches the same conclusion from the ops side: simpler patterns often deliver 80% of the savings at 20% of the overhead, and sophistication pays only after ground-truth validation and config management exist. https://seldondance.substack.com/p/production-ready-llm-routing #LLMRouting #AIInfrastructure

3 more comments...

Ready for more?