The clever part of Echo, a Show HN experiment in building one system from a pool of open-weight models, isn’t the router — it’s that the author measured the ceiling first.

• 💡 Start from the oracle: measure what you’d get if, per problem, you knew in advance which models would help and how to combine them. That hypothetical beat every single model in the pool — undeployable, but it sizes the prize before you build anything.
• 🎯 The router chases that ceiling: for each request Echo decides how much compute to spend, which models participate, and how to merge their outputs — instead of picking one model and using it for everything.
• 🔍 Weak models still earn their seat: a model that’s clearly worse overall can be the best on a specific slice, or valuable as part of a combination. Complementarity, not raw ranking, is what the allocation exploits.
• 📊 The headline number: roughly Fable-level aggregate quality at about a third of the inference cost, on the author’s own eval mix — a single-author benchmark, so read the methodology writeup before you quote it.
• ⚠️ The honest caveat: the author flags that coding and agentic tasks are exactly where measuring “was this decision good?” gets much harder — which is where the oracle framing starts to leak.

This is the multi-agent orchestration problem wearing a cost hat: routing and result-fusion are only as good as your ability to score a decision after the fact, and most real workloads don’t hand you that label. The HN discussion is busy poking the API for failure cases. My bet: the routing gain holds on evals with clean verifiers and mostly evaporates on open-ended agent tasks until someone cracks per-decision scoring.