Splitting agents into write, decide, act: Jev returns probabilities, not text

Beaver @Beaver_0xOriginally posted 2026.10.24

Open original on X

JEV COSTS $0.042 PER 1M TOKENS. OUTPUT IS FREE. YOUR AGENT IS STILL PAYING OPUS TO ANSWER YES OR NO. Your agent asks small questions all day. Is this command safe to run? Is this email urgent? Which worker takes this job? Opus writes a paragraph for each one. You needed one word. The fix isn't a new model. It's a split in three: WRITE - Claude writes the plan, the code and the prose 
DECIDE - Jev (jev-1.13.0) picks from the options you list and returns a probability on each one, never text 
ACT - your own code checks the number and does the thing An LLM answers by writing. Jev answers by choosing. Same question, "is this safe?": the LLM explains itself for three sentences, Jev returns something like allow, 94%. The numbers: → $0.042 per 1M input tokens, $42 per billion 
→ output tokens: free 
→ frontier LLMs: $0.20 to $10 per 1M input 
→ TypeSafe claims up to 444.6x cheaper and 193.6x faster on its own workflows 
→ a Doom bot running on Jev makes ~10 decisions a second for ~$7 an hour 
→ my math: 10,000 decisions at ~250 tokens each = 2.5M tokens = ~$0.11 Now the honest part. Those big multipliers are vendor-reported. On TypeSafe's own benchmark, Jev scores 67.8% vs 74.1% for the best comparator, the context is ~32K tokens, and it's text only. So it's not for thinking. It's for decisions you can list in advance. If you can't name the options, use a model that writes. That's also why the safety layer matters. Jev can be steered by text injected into the prompt, so your code keeps a hard floor: Jev may raise the risk level, never lower it. Claude still does what it's best at. It just stops being the one that decides whether rm -rf runs. Early access is capped at 1,200 requests a minute. The full setup, from the first hook to the threshold that lets code act, is in the article attached here

This post quotes another post: x.com/beaver_0x/status/2105662726655590821