Theo: model routers can't gauge task complexity from prompts alone; escalation-based routing proposed
nbaschez · x · 2026-09-26
Theo argues that model routers like Jev are "system one" models—they categorize but don't reason. Task complexity is nearly impossible to judge from a prompt alone: you can't tell if "port this to Rust" targets a 100-line file or a million-line codebase, and the router has zero context on the agent's tools and environment.
Replying, nbaschez suggests the ideal router starts with a cheap model that learns to escalate as needed, with a stronger model acting as a checker that can also escalate.
More from coding & agent
- Sentry CEO: RAG isn't dead — context fetching is core to any good agent — zeeg · 2026-09-26
- Adam Dorr: agents looping scripts is easy; solving real problems remains marginal — adam_dorr · 2026-09-26
- Grok agent in a Tesla fetched a work parking pass end-to-end, unprompted — Baconbrix · 2026-09-26
- In an AI-first world, your own understanding of the project is state to preserve — brandon_xyzw · 2026-09-26
- Railway launches free no-account VMs (59 min) and OpenCode cloud agents — jasonkneen · 2026-09-26
- 575 PRs later: how the Astra agent untangled a sync-SQLite bottleneck in OC — soumitrashukla9 · 2026-09-26