Jev router cuts agent harness costs in half in 32-call benchmark test
dair_ai · x · 2026-09-26
- Jev from TypeSafe AI is a tiny System One model built solely for harness micro-decisions: it takes a situation plus a few questions and returns numbers, never text — cheap enough to run on every agent step.
- It's now packaged as typesafe/jev-router on OpenRouter: send the conversation and it picks the model and reasoning effort per request instead of manual selection.
- DAIR.AI tested it on a support agent built with the Pi SDK: same 8 cases, 32 real calls, vs a fixed GPT-6 Sol baseline. Both got every case right; the router cost less than half ($0.008 vs $0.018) with lower median latency.
- A companion tutorial covers building the harness with Pi SDK, using Jev to pick models, block risky tool calls, and self-check answers.
More from coding & agent
- PHASEONE agent's redacted "64H" hints the real IP is real-time budgets, not token budgets — 1a3orn · 2026-09-27
- Yacine returns to work: frontier models let one person red-team like entire teams used to — yacineMTB · 2026-09-27
- Hindsight, a ~30K-star project, gives coding agents long-term memory from git history — techNmak · 2026-09-27
- dhh: programmers who can beat a pair of collaborating frontier models round to zero — suchenzang · 2026-09-27
- Agent fixed what its reviewer flagged—then the next review passed the new bug it created — Fit-Ant2238 · 2026-09-27
- Meta Muse hits 3.4M downloads; an indie dev is building the on-device, privacy-first alternative — jonathancheckwise · 2026-09-27