I’ve seen this name in other articles too—I’m curious what it’s about. https://t.co/7qiEKzo5Yd Jev AI has changed the common sense of AI agents Julian Goldie SEO #AISummary #Claude #Codex Jev: Transforming AI Agent Decision-Making and Automation 🔳 Jev’s Key Features Jev is a System 1-type AI model specialized not in text generation, but in decision-making tasks such as selection, evaluation, and judgment—designed to process large volumes of decisions quickly and at low cost. 🔳 Role Division Compared to Traditional Models While large models handle research, planning, and text generation, Jev handles brief, high-frequency decisions: selecting the next tool to use, determining whether to continue processing, or assessing whether human review is needed. 🔳 Speed and Cost The video cites figures showing Jev is 20x to 200x faster and 40x to 400x cheaper than standard AI models, significantly reducing the need to offload numerous small decisions to high-performance models. 🔳 Three Types of Judgments Jev supports three judgment formats: Choice (selecting from multiple options), Score (evaluating against defined criteria), and Probability-based Yes/No classification—each with an accompanying confidence score. 🔳 Automation Based on Confidence Levels You can configure rules to auto-process decisions with high confidence and route low-confidence decisions to humans—allowing systems to account not only for potential errors but also for uncertainty in judgment. 🔳 Parallel Processing of Multiple Judgments Multiple questions about the same context can be submitted simultaneously to evaluate next steps, urgency, and need for approval—all in parallel—reducing wait times and costs within AI agents. 🔳 Research Paper Classification Example After summarizing AI research papers, the developer fed Jev titles, summaries, and 24 candidate topics for classification. In the video, processing 18 papers cost a total of 8 cents, with a median processing time of 256 milliseconds per paper. 🔳 Email Sorting Example In a test sorting 500 emails, Jev categorized them into options like “Reply,” “Further Research,” “Hold,” or “Human Review,” completing the task in seconds at a total cost of 3.5 cents. 🔳 Lead Evaluation Example Evaluating 700 leads with individual messages, Jev predicted response likelihood and confidence levels in 40 seconds, identifying mismatches between leads and messages—all for just 9 cents. 🔳 Browser Agent Integration In browser agents, Jev is used to select the next action each time the page changes, while text input is handled by a separate generative model. The video shows browser command counts reduced from a median of 1,092 to 101, cutting task time by up to 25%. ListItemIcon Flight Search Note The example showed flight discovery in 7 seconds at under 0.5 cents—but no booking was executed. The 7-second measurement began only after the initial page load. ListItemIcon Internal Link Optimization For a site with 586 pages, Jev placed 584 internal links in 45.1 seconds. For the 139 pages lacking suitable destinations, no links were added—at a total cost of 21 cents. ListItemIcon Model Routing Jev can route tasks between inexpensive, fast models and expensive, high-performance models based on task requirements. The video illustrates integration with LangChain: selecting the cheapest viable model for each task. ListItemIcon Safety Check for Tool Usage Before an AI agent executes a tool, Jev can assess risk: approving execution if safe, or halting it and triggering human review if risky. ListItemIcon Context Reduction Jev evaluates the importance of past tool calls in agent history and removes unnecessary information—in one example, reducing context from ~1 million tokens to ~86,000 tokens in under one second. ListItemIcon Concerns About Context Reduction Critics note that individually evaluating and deleting history entries may lose critical reasoning or context over long tasks. Alternatives like summarization instead of deletion are suggested as potential improvements. ListItemIcon Jev’s Limitations Jev does not generate text, code, or research findings—it is designed solely for decision-making. High confidence does not guarantee correctness; actual outcomes (e.g., file saves) must still be verified separately. ListItemIcon Importance of Input Design Judgment intent cannot be conveyed by question field names alone—specific phrasing of questions and options is required. Poorly defined context or insufficient information will degrade judgment quality. ListItemIcon Pricing Structure The video states input costs are approximately 4 cents per 1 million tokens; output currently incurs no charge. For 10,000 judgments using 1,000 tokens each, estimated cost is roughly 42 cents. ListItemIcon What Matters Most: Total Task Completion Cost🔳 The key is task completion cost: Even if individual decisions are inexpensive, unnecessary processing by agents due to incorrect routing increases overall costs; therefore, it’s essential to evaluate the total cost to complete a task, not just the per-decision cost. 🔳 Structural changes in AI agents: The video explains that whereas previously a single model handled both text generation and decision-making, a new division of labor has emerged—large models now handle research and generation, Jev manages selection, evaluation, approval, and escalation, and code executes actual operations. 🔳 Current suitable applications: The integrated functionality is still in an experimental stage; it is more practical to begin implementing it in simple, well-defined decision-making tasks that humans repeatedly perform, such as email classification, lead evaluation, routing, and flagging. 🔳 Video conclusion: The central argument of the video is that separating text generation from decision-making—and delegating repetitive, small-scale decisions to fast, low-cost models like Jev—can improve the overall speed, cost efficiency, and scope of automation for AI agents.