System One decision models reshape custom agent harnesses as CLM hits 9x Jev speed
_AustinCalvert_ · x · 2026-09-26
A discussion sparked by DAIR.AI's Omar argues that a new wave of "System One" decision models changes the economics of building custom agent harnesses. These small models handle the harness's frequent micro-decisions — which model to route to, whether a tool call is safe, whether an answer is good enough — without a full LLM call each time.
Key points:
- After Jev (TypeSafe AI, trained with RLCD, returns typed decisions with probabilities), a new Contrastive Language Model (CLM) embeds situations and candidate actions and ranks by similarity — 9x faster than Jev and seemingly a better verifier on long-horizon tasks
- Why own your harness: control of your intelligence stack, System One + System Two combos beat either alone, richer proactive agent experiences, more value per token
- A DAIR.AI Academy lab shows how to build a custom harness on the Pi SDK using Jev for model routing, risky tool-call blocking, and answer checking
- Author expects frontier labs to push decision models further with better customization tooling
More from coding & agent
- Claude Opus 5.5 generates a full launch video — animation, music, voiceover — in 20 minutes — cedric_chee · 2026-09-26
- OpenRouter launches typesafe/jev-router, a cache-aware router that picks models per request — alexcovo_eth · 2026-09-26
- Data labeling firm predicts enterprise evals will become core IP as revenue shifts to Fortune 1000 — soumitrashukla9 · 2026-09-26
- Task-Specific RL Won't Yield General AI Agents, Researcher Argues — xuanalogue · 2026-09-26
- asc builds upload now prints build IDs and surfaces Apple's real errors — rudrank · 2026-09-26
- asc CLI pre-checks emojis locally, naming each character Apple would reject — rudrank · 2026-09-26