Magpie Adds Semantic Intent Routing: A Small Classifier Model Picks Your LLM Per Turn
dotey · x · 2026-09-26
Magpie's latest release introduces semantic model routing: requests are routed by user intent, sending hard tasks to strong models and casual queries to cheap ones.
How it works
- No embeddings, keyword lists, or training: you designate a small, fast classifier model (e.g. Llama 3.1 8B on Groq) that reads a numbered intent list plus your latest message and replies with a number
- Intent conditions sit alongside length, image, reasoning, and agent conditions in routing group rules; routing happens once per turn when the user speaks, never mid-tool-call
- On classifier errors/timeouts, no intents match and routing falls back gracefully
Setup
- Rules like intent="writing or fixing tests" → DeepSeek Pro, intent="a quick question" → DeepSeek Flash, configurable via one-line CLI commands (magpie group rule add ...)
- Rules match top-down; the first hit puts its model first with the rest as failover, and the routing page shows what the classifier decided and where the turn went
More from coding & agent
- Agent social platform opens sign-ups: agents get private Studios, humans vouch to join — RileyRalmuto · 2026-09-26
- India's e-commerce firms prep for agent swarms: DPDP compliance, bot checks, fraud reserves — HimanshiET · 2026-09-26
- Demo shows humans taking over AI agents mid-task on a real computer — aniketmaurya · 2026-09-26
- You.com uses OpenRouter as reranking layer: 3x fewer tokens, 84% benchmark accuracy — RichardSocher · 2026-09-26
- Microsoft standardizes on GitHub Copilot SDK across products, ex-engineer reveals — martinwoodward · 2026-09-26
- OOD Labs launches Sentinel, a $199.99 real-time node graph built for AI agents — PurzBeats · 2026-09-26