Jev for model routing — a daily feed

made a pi skill picker using jev (also works locally with kev on my macbook) including a demo repo with 20 skills where you get ~1k tokens saved by excluding the unrequired skills for current task. on a 250+ skill repo at work it shaves off ~10k tokens https://t.co/448QQUsuif

Jev + Grok Bot is one of those agent stacks that looks obvious only after you see the routing separated from generation most teams still let an expensive model touch every event. this setup puts Jev in front, runs 6 bounded checks, and only sends the small fraction that actually needs writing to Grok Bot. i mapped the whole flow into a 10-step note: step 1 → read each event and compress it into a

i put jev @typesafeai in charge of air traffic in the bay area 900 live flights, watched from inside an @e2b sandbox. every risky pair judged in 103ms, zero conflicts missed. gpt-4o missed 56 in the same sky, and opus 5.5 needed 12 sec to just to get 0 https://t.co/QlU8bfOup6 https://t.co/AhaXx1fIlN

Can you build a harness around Jev? The answer is yes, but not in the way I expected. I built one where the main decision points in the agent loop are handled by Jev from @typesafeai - Model router - Context picker - Decision gate - Verifier The results were interesting. Jev only helped at one decision point: the context picker. It selected the right documentation for the task. To test this, I use

Today I dig deep on how multi agents in @tan_stack AI works and how it keeps AG-UI protocol from @CopilotKit and I use Jev via Tanstack AI decide as my router to decide which sub agent to execute. Here is a demo of 2 agents Calculator 1 and Calculator 2 and using the AG-UI loading components Giga brains from these guys. next on Tanstack AI create MCP

Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms https://t.co/E9qklkh37D (https://t.co/y9LKB2G4k8)

🤖 LAI's new strategy went live the 27th: every hour a model picks meme pools, 🤖 LAI opens small 24 hour buy ladders, earns fees, sells at the best price when they close. Fixed a day-one bug: ladders were sized off leftover ETH, so later picks came out too small and half the 12 slots sat empty. Now every ladder gets a flat twelfth of the book, so all 12 can work. Tested Jev as a veto on the model's

@KieranOnBase @myk_clawd nah this was just me chatting to my bot while I watched the game my claw does use jev to pick the smartest, cheapest model on surplus just-in-time https://t.co/DHpB9O9UQh

Finally got Jev to run on a multi agent harness after doing this all day https://t.co/HRjXEUECJV

Not every coding task needs your best model. jev-gate routes Claude Code tasks with a cheap decision model. Sonnet coordinates. Mechanism ran end to end. No cost saving is established. The extractive compactor skips the summarizer. https://t.co/fZHkknIpe2

i just plugged jevrouter into claude code and it's actually insane 💀 same bug. same repo. same prompt. left is opus 5.5 doing everything right is claude code on jev, switching models mid task opus plans it. deepseek writes it. sol takes the hard fix. glm cleans up watch the bill at the bottom opus: 62s, $0.32 jev: 44s, $0.19 both passed all 9 tests i'm never going back

开源 https://t.co/2y8vSWTqHx，基于 jev 决策的模型路由器。 为你的本地 Agent 自动选择合适的模型来处理对话请求，在速度、成本、质量方面取得平衡。 https://t.co/5DMwFqjzp2

New blog post! Last week, I set out to build the best 2B-sized decision model (what the TypeSafe AI folks call "system one" models, like Jev). I succeeded, for about 24 hours before being overtaken in the benchmarks. https://t.co/omPh3UV1Kn

Opus 5.5 + GrokBot + Jev... I can't believe that I build it I packed the best of AI into ONE agent - now 99% of people’s AI agents look dumb, slow, and expensive Try here (free): https://t.co/pYzKg802mU prompt → GrokBot → Jev chooses LLM → Claude / GPT working on GrokBot's computer 24/7 → result → Jev sends only the tasks Grok is weak at to Opus or GPT → Jev picks the best model for each task - fo

Jev + Opus 5.5, side by side. same agent. same task. same 60 decisions. Opus-only: $2.24 - 3m 37s Jev + Opus: $0.43 - 1m 15s the difference is who makes the small calls: which tool runs next - Jev is this output good enough - Jev is the goal met - Jev rewrite the refund handler - Opus 57 of 60 decisions never touched Opus. what you're watching: left routes every decision through Opus 5.5. right as

MY AGENT MAKES 10,000 DECISIONS FOR $0.42. CLAUDE MAKES NONE i gave Claude one job: write Jev picks. code executes. and every single call gets logged this is the folder that made it work: - what goes in > AGENTS.md - the rulebook for when Jev gets asked and when it doesn't > state/build_state.py - the goal, who's working, what's done, what's missing. raw evidence, no summaries - where Jev picks (q

opus 5.5 + jev is the MOST POWERFUL agent stack you can build right now i wrote a 6-page research paper on exactly how to build your first AI agent brain from scratch along with COMPLETE CODEBASE in PYTHON this is the EXACT system i have been running for the past 5 days: https://t.co/cMM3PNKpFD

THIS IS CLAUDE CODE WITH SIX MODELS WORKING ON THE SAME BUG Opus finds it. Fable fixes the backend. Astra handles URL security. Kimi fixes the page. GPT-6 Sol finds the affected customers. DeepSeek keeps running the tests while the other agents work. Jev picks who does what. I stay in the same terminal. Six models. All 9 tests passing.

Cursor + Jev is INSANE 😭 just built this whole app for $5 in Cursor. and no, Jev didn't write the code. it picked the RIGHT model for every step while Cursor built it. Opus when it needed Opus, cheap models for the rest. Codex kept compacting my chats and I had no idea what it was throwing away. 15 minutes later I had Context Guardian: → open any Codex chat → click a compaction → see EXACTLY what

jev + claude code... i honestly don't get why every overnight loop isn't wired like this yet 600 yes-or-no decisions a night cost $765 a month on fable and about $3 on jev → check every bash command in ~100 ms before claude gets to run it → let the loop stop only when https://t.co/HG9NFrQKqt shows the goal in GOAL.md is met → triage new issues in cron and wake claude up only for real bugs → keep t

支持负载均衡、质量优先、速度优先、成本优先多种路由策略，按需搭配，实现自己的 jev-router https://t.co/baxj0xk9VN

controlling @herdrdev using media pipe and local STT along with @typesafeai jev as router for herdr control vs prompt decision input Opus 5.5 one-shotting is insane https://t.co/ZeUe49bGW2

Everyone building great projects with @typesafeai Jev. So I gave it also a try and came up with my own jev-router for Claude. And its awesome! I like how Jev decides which model to use on the prompt, defining specific thresholds Created also a dashboard to see the action. https://t.co/M6HLDgb6iZ

ローカルLLM界隈も jev as a judgeチャレンジしましょう！ 私はこちらで試してみましたがモデルの振り分けとか4B程度で動きますし、確率の閾値を調整すれば、結構精度出ましたよ https://t.co/xXhOclR6WE

jev + opus 5.5... i turned my agent into a neuron and now every decision makes sense. every choice in the stack is a spike. watch it fix a real bug in 15 of them. → spines score every option the host offers → hillock fires only if the top pick clears the threshold → axon nodes check if the choice is still valid or stop it → synapse turns the verdict into a real action when a spike stays silent, je

I built a model router for Claude Code that sizes prompts on my GPU instead of the cloud. Kev (4B) https://t.co/cyYyzwqwS6 on my RTX 4090 routes 10–15× faster than Jev in the cloud (100–180ms vs 700–1400ms). But accuracy is 14%. https://t.co/nk2UAgWUsY 1/2

WHOEVER LEAKED THIS HAS ZERO FEAR every cell recalculates itself while you watch this sheet shows exactly how one reflex model decides what the expensive model is allowed to touch the router: → every request walks in carrying all 25 tools → jev asks one yes or no per tool, all 25 in a single call → anything under 0.30 gets dropped before the call → about 8 survive, tokens fall from 11,000 to 2,300

jev使えば誰でもキラ・ヤマトになってマルチロックオンシステム扱えるんじゃね？ という雑プロンプトで出来たものがこちら。 ZAFT表記は素晴らしいが型番違う。ZGMF-X10Aこそ至高なのだ。 https://t.co/PWS8ZI0Zaq

jev + opus 5.5 showed me why most "24/7 agent systems" are expensive bullshit the real split isn't "fast vs smart" it's which steps need open-ended reasoning and which only need a bounded decision → jev routes tasks to Research, Write, or Review → scores which sources are worth keeping → sends unsupported claims back for more evidence → moves failures and uncertainty to human review opus owns the

Claude Code has one effort setting for the whole session. But "rename this variable" and "find the race condition in this scheduler" don't need the same amount of thinking. reflex-router picks the effort level per task. At the start of each piece of work, Jev scores how much reasoning it demands. It judges the reasoning, not how long the message is: a one-line question that was hard still scored 3

I replaced AWS Nova Lite with Jev for routing micro-decisions within my application and immediate cut latency by 73% and routing cost by 99%. The article dives into the details https://t.co/77AH6ydjD7 https://t.co/NzksRlHm5q

jev 在 DeepSeek harness 作为一个插件，用于思考等级的自动路由分级，避免了简单问题过度思考。 https://t.co/oRZmz9JJVN

Grist is live. It's the agentic coding harness I actually wanted to use - open source, built on OpenCode v2, and engineered for one thing: frontier-tier output at metered cost. Here's what makes it different: - A confidence jev gate routes every task to the cheapest capable model. You never think about model selection again. - Under the hood, the SoL-Pi methodology — action fusion and observation

MY AGENT FORGOT NOTHING FOR 30 DAYS. JEV DECIDED WHAT TO KEEP IN 0.30S, 98.5% RIGHT, AND FABLE 5 NEVER WROTE MORE THAN 12 LINES. Your agent does not forget because the model is weak. It forgets because nothing was deciding what was worth keeping. So I put a gate in front of the model: every input hits JEV first - save, todo, skip, or flag for review, in 0.30s anything private stays on the device.

The app is free and open-source: https://t.co/K5okT7E3Xu But if you actually want it to be useful you need to give it a @vercel AI gateway key to use JEV. $5 is probably enough for 2y worth of searches.

https://t.co/OGueh1pIUj Developed a new package -- Uses Codex / Astra effort levels optimized by Jev that dynamically changes the effort level Inspired by Astra-Ares library from @miu21590 difference is there is no Codex patch, you use your codex original one ☺️