Jev · @Bober_smart
Jev is driving a massive cost revolution in AI: executing 10,000 decisions now costs $0.42 instead of $300 Here is a breakdown of what this tech actually is, h
Jev is driving a massive cost revolution in AI: executing 10,000 decisions now costs $0.42 instead of $300 Here is a breakdown of what this tech actually is, how to implement it, and how to spot overpriced API calls inside your existing agents Phase 1: Audit and Core Mechanics Identify low-hanging fruit: Check your agent's current prompts. If the potential choices are fixed, a human could make the call instantly, and high frequency makes speed essential, remove it from the LLM immediately Pure structured data: Pass the system state alongside strongly typed queries to retrieve exact probabilities. Skip text generation entirely to eliminate unexpected outputs and parsing issues The 3 decision primitives: 1. Choice: Selects a target option out of up to 255 predefined values 2. Score: Maps current state onto a custom numerical scale 3. Noul: Outputs the exact probability of a simple "yes" outcome Phase 2: Quick Start and Testing Environment setup: Initialize typesafe-sdk in a clean virtual environment via pip, add your credentials to environment variables, and pull the official skill with npx if an AI agent is writing your codebase Sandbox verification: Before pushing to production, test a single decision in the Playground. Feed a sample ticket as state, throw in a Noul and a Choice call, tweak a field, and observe the resulting distribution shift Phase 3: Scaling and Architecture Best Practices Parallel evaluation: Batch multiple questions into a single request. Evaluate everything concurrently, leaving the branching logic entirely to your code Production decision gates: Let traditional code aggregate file counts and test statuses, while Jev handles final actions (ship, return, or escalate) alongside safety checks Optimized batching: Group up to 8 queries per request without hitches in latency. If you require real-time updates mid-execution, spin up a secondary request Dynamic state menus: Filter active worker availability and source availability at the code level before handing options over to Jev Safety harness: Put middleware in place to validate tool execution and deploy a smart router that picks the most affordable model capable of resolving the task Real-World Benchmark Examples Browser Use executed a live flight search in 7 seconds for just $0.0039 A developer categorized 1,018 academic papers for a total of $0.08 500 emails were fully sorted for 3.5 cents Golden Rule: If deterministic code can solve a problem, never hand it over to an LLM Take a look at your last agent execution log and count how many decisions never required a single word of text generation Full implementation details and code snippets are available in the attached article below
这条是引用着另一条写的：x.com/bober_smart/status/2103581364134256863