Routing 1000+ video models: filter first, score second, never trust the agent to choose
Vaciuum1 · reddit · 2026-10-10
A backend dev of an AI video agent shares routing lessons: letting the agent pick from 1000+ models led to confident but wrong picks (no reference-image support, wrong aspect ratio). Their fix: (1) hard-filter models that can't handle the user's inputs before the agent sees them; (2) score the rest (release date, description, user preferences like 'always use seedance'); (3) let the agent pick only from a vetted shortlist; (4) show model + cost estimate with nothing running until user approval; (5) pull model schemas every 10 min so provider changes don't silently break jobs. Worst bug: one unreadable schema wiped the entire catalog. They ask: learned picker vs rules + scoring?
Related event: Dev Shares Routing Tips for 1000+ AI Video Models(2 posts)→
More from coding & agent
- Indie dev builds a GT5-grade browser racing sim in ThreeJS, tuned to IMSA GT3 rulebook — AIandDesign · 2026-10-10
- Agent-built system hits 2,242 tok/s on AMD MI300As, 2.33× faster than SGLang in 105 hours — bariskasikci · 2026-10-10
- Seroter's Daily Reads: Agent-Friendly APIs, LLM Judges, and Ambient Quality Agents — rseroter · 2026-10-10
- Steve Yegge: Claude just reads the binary when it wants to know how closed-source Rex works — Steve_Yegge · 2026-10-10
- YC hosts an agent-first hackathon in San Francisco, sponsored by Supabase — ycombinator · 2026-10-10
- Running a fleet of long-horizon LLM agents: 9 keep-alive patterns and 5 unsolved problems — milkygirl21 · 2026-10-10