Price: $0.05527 0.2262%
Market Cap: $8.67B 0.415%
Volume (24h): 401.43M 0%
Dominance: 0.415%
Price: $0.05527 0.2262%
Market Cap: $8.67B 0.415%
Volume (24h): 401.43M 0%
Dominance: 0.415% 0.415%
  • Price: $0.05527 0.2262%
  • Market Cap: 8.67B 0.415%
  • Volume (24h): 401.43M 0%
  • Dominance: 0.415% 0.415%
  • Price: $0.05527 0.2262%
Home > 视频 > The Code Is Cheap, Who Approves It? Jev, OpenAI Decisions API And How PM Sets Thresholds For AI Agents

The Code Is Cheap, Who Approves It? Jev, OpenAI Decisions API And How PM Sets Thresholds For AI Agents

Release: 2026/10/26 17:06 Reading: 0

Original author:Tonyrising

Original source:https://www.youtube.com/embed/__5s6FGeAk4

A pull request written by an AI agent only takes a few minutes. But according to LinearB's report on more than 8 million pull requests in 2026, one in four pull requests written by agents had to wait more than 17 hours for someone to approve. With human-written code, that number is about 3.5 hours. Code is very cheap now. There are still only so many people reviewing. And last September, the market began to respond in a new way: letting AI do the reviewing. This episode goes through three questions: what's going on, why do agents need "fast brains", and what do PMs, POs, devs, and testers need to do. 🎯 In this episode: • Cheap to make, approval becomes a bottleneck; Individual agents running in the background (dots, Muse, Claude Duet) must request permission all day. • Jev (TypeSafe AI, September 2026): "System One" model does not write text, only returns choices, scores, yes/no with certainty level. Nearly 13% of paid teams on Vercel tried it on the first day. • OpenAI Decisions API: launched about three weeks later, OpenAI cites Jev as inspiration. • Price with date: Jev 0.042 USD / 1 million input tokens (Fortune, October 8, 2026); Decisions API 0.10 USD / 1 million input tokens when open to everyone (October 22, 2026). Both do not count output tokens. • Metaphor: triage nurses in the emergency room are the decision model; The specialist is a major LLM. ⚡ WHY THIS TYPE OF MODEL WAS BORN • Cost: pay for a label, not the whole paragraph. • Speed: up to 10 times faster (according to OpenAI). • Control: output has a fixed form, with a level of certainty, so thresholds can be set, logged, tested. • Objection: fast models can still be certain but wrong. ✅ PM, PO: 4 THINGS 1. "Threshold register" for running agents: decision point, threshold, never automatic tasks, threshold approver, "sure but wrong" cases/week. 2. Prioritize safety in the wrong dangerous direction: if not sure, rank it to a higher level; Irreversible things always depend on the person. 3. Try classifying pull requests: low risk goes to automatic flow, high risk goes to the best reviewer; Measure pre- and post-approval wait times. 4. Calculate agent costs by number of decisions, not just number of answers. 🛠️ DEV, TESTER: 3 THINGS 1. The test data set has the correct answer; Measure to see if the model says it's 90% sure and is about 90% correct. 2. Log each decision with its level of certainty; Separately monitor "certain but wrong" cases. 3. Where the user needs an explanation, keep the LLM. ➕ With outsourcing: agent quote separates two lines, thinking cost and decision cost. 💡 Do it this week: open the threshold register with the first three decisions of the project. ⏱️ TIMESTAMPS 00:00 Introduction: code written in a few minutes, waiting for approval 17 hours 00:24 The code is very cheap, the number of people reviewing is still the same 00:37 Hello channel and note about the source 01:05 Today's topic: 3 questions 01:28 Question 1: Cheap to make, approval becomes a bottleneck 01:44 Personal agent running in the background: Muse, dots, Claude Duet 02:10 Jev: the model doesn't write, just decides 02:48 OpenAI Decisions API, three weeks later 03:02 Price with release date 03:20 The expensive thing is who approves it, and can it be approved in time 03:30 Question 2: Triage nursing metaphor in the emergency room 04:16 Reason 1: cost 04:27 Reason 2: speed 04:40 Reason 3: controllable 04:52 Vercel uses Jev to automatically review the agent's commands 05:12 Comment: fast can still be sure but wrong 05:33 How much certainty can you do yourself? 05:44 Question 3: What do PMs, POs and people in the project need to do? 05:57 Task 1: threshold register 06:19 Task 2: if you're not sure, raise it to a higher level 06:32 Task 3: try to classify pull requests 07:04 Dev, tester: 3 tasks 07:28 Outsourcing corner 07:41 Do it this week 08:01 Does the agent know when he's not sure? 08:12 In short: you approve the threshold 08:35 Question for you 💬 QUESTION FOR YOU In your projects, how many places are using AI to write an entire paragraph, just to get one word "yes" or "no"? Tell me about it in the comments section. 📺 See more: • One Week, Six Security Incidents: AI API Keys Are Being Hunted, What Should PMs and Teams Do? (episode 18) https://youtu.be/QLktJ9EUydI • Google Lets AI Agent As "Colleague": When The Job Is Finished, Who Lets It Leave? (episode 17) https://youtu.be/aarC5IjpdfY 📚 Source: • LinearB: 2026 Software Engineering Benchmarks (8.1 million pull requests) • Fortune (October 8, 2026): Jev, an AI for making quick decisions, has been a viral hit • Vercel (September 19, 2026): Auto-approve tool calls with eve and Jev • Unite.AI, Crypto Briefing (October 22, 2026): OpenAI Decisions API generally available • MacRumors (October 21, 2026): Anthropic launches Claude Duet • The Next Web: OpenAI launches dots ⚠️ Information in the video is compiled from public sources at the time of making the video. The number 17 hours is the wait time at P75 (1 in 4 pull requests wait longer), according to LinearB. The numbers 200x and 400x announced by TypeSafe themselves are only correct for the classification problem. Comments and solutions are the host's personal perspective; Each project has a context, please refer to and adjust. 🎙️ The Future of AI, with Tony Rising: AI news and analysis for Project Managers, Product Owners and IT workers. 👉 Subscribe and turn on the bell so you don't miss new episodes. #AIAgent #AIGovernance #CodeReview #ProjectManagement #TheFutureOfAI

Selected Topics

  • Dogecoin whale activity
    Dogecoin whale activity
    Get the latest insights into Dogecoin whale activities with our comprehensive analysis. Discover trends, patterns, and the impact of these whales on the Dogecoin market. Stay informed with our expert analysis and stay ahead in your cryptocurrency journey.
  • Dogecoin Mining
    Dogecoin Mining
    Dogecoin mining is the process of adding new blocks of transactions to the Dogecoin blockchain. Miners are rewarded with new Dogecoin for their work. This topic provides articles related to Dogecoin mining, including how to mine Dogecoin, the best mining hardware and software, and the profitability of Dogecoin mining.
  • Spacex Starship Launch
    Spacex Starship Launch
    This topic provides articles related to SpaceX Starship launches, including launch dates, mission details, and launch status. Stay up to date on the latest SpaceX Starship launches with this informative and comprehensive resource.
  • King of Memes: Dogecoin
    King of Memes: Dogecoin
    This topic provides articles related to the most popular memes, including "The King of Memes: Dogecoin." Memecoin has become a dominant player in the crypto space. These digital assets are popular for a variety of reasons. They drive the most innovative aspects of blockchain.