Tech news picked on 2026-11-07
Kiro Crew Ultra Mode splits big jobs across AI chats and manages them on one screen
- Ultra Mode in the open-source tool Kiro Crew splits a single goal into multiple tasks, assigns each to a separate AI chat, and brings progress and approval requests together on one board.
- A conductor AI sets completion criteria for each task and writes them in a ledger (task log). Task AIs only report results and cannot change their own completion criteria or mark themselves as passed. Records remain after a restart.
- It arrived as a feature preview in 0.12.0 and is off by default. By default, each task assignment requires approval, and the default number of tasks opened simultaneously per goal is 6. The author derived this figure by reading the code.
- In the author's server inspection test, approval was required to assign tasks, run commands, check completion criteria, and close chats. When a task AI reports that it is finished, the board shows it as complete, but the ledger entry remains open until the conductor AI checks and closes it.
- Thinking Machines Lab, founded by former OpenAI executive Mira Murati, has released a 2-generation AI model. It is designed so that companies can further train the model on their own data and adapt it to their own work.
- For now, the weights (the learned contents of the model) are closed, and it can be used only through the company's fine-tuning service, Tinker. The company said it will release the model under the Apache 2.0 license next week, after completing an external safety review.
- According to the company, the architecture activates only 17 billion parameters out of a total of 594 billion parameters, and it supports image input. Its Artificial Analysis Intelligence Index score is 31, which the company says far exceeds its previous model.
- In the coding evaluations of the open model comparison group chosen by the company, it ranked number 2 in 4 of 6 evaluations and never ranked number 1. On the GPU code optimization evaluation PMPP-Hard, it fell short of GPT-6 Astra and Claude Opus 5.5.
Preview a remote server screen in VS Code without port forwarding
- In VS Code 1.145, you can view an AI-generated web app running on another computer directly inside the editor on your own computer. No separate port forwarding setup is needed to view it.
- The author connected to an Azure virtual machine over SSH and opened a sign-up screen. They said an AI agent controlled the browser to check input errors, successful registration, and duplicate registration.
- You must enable workbench.browser.enableRemoteProxy in settings. Both the user's computer and the remote host must be running supported versions, and cloud sessions are not supported.
- You can share a logged-in tab that you opened yourself with the agent, so take care not to share tabs displaying confidential information.
Should we understand code written by AI? A developer's worry about speed and quality
- This is a piece by a developer who has believed that people should understand and modify the output themselves, and who is now rethinking their way of working as generative AI increasingly builds and fixes programs.
- The author feels that the quality difference between people who submit AI-generated code as is and people who submit it after understanding it is not clearly visible from the outside. This is the author's personal experience and opinion.
- The author asks who would be rewarded: a person who produces 3 results worth 70 points in the same time, or a person who produces 1 worth 90 points. The scores are a hypothetical example to explain the choice between speed and quality.
- The author sees a trend of leaving everything to AI, from deciding which features a program should have to development, testing and maintenance. The author points out that once a product is released, development may be judged a success regardless of usability.
- Wiredy uses AI to recommend influencers suited to a brand's products and shows projected sales volume in advance. Once a sale is arranged, the system handles everything from setting up an online store to orders, delivery, customer service and settlements.
- According to the company, it was trained on more than 20,000 actual group-buy transactions accumulated over 8 years. Its projected revenue error rate is reportedly about 15%, lower than manual forecasts by managers with at least 5 years of experience (about 20%).
- In 2025, the service's transaction volume was 33.6 billion won, and revenue was 29.2 billion won. It officially opened a dedicated online store in May 2026 and began charging for it in July. It has worked with 2,000 brands so far.
- It has secured a database of about 130,000 Japanese influencers and is running campaigns with Korean brands. It plans to launch an AI agent for Japan in 2027, then expand to the US, Taiwan and Singapore.
AI customer support design: leave document search to AI and contract decisions to DB rules
- This is a manga-style introductory resource written by Yuji Arakawa of Oracle Japan. A new engineer designs a support system that finds the manual matching an inquiry and checks contract information to see whether an on-site visit is possible.
- Document search finds items similar in meaning to the question, but first narrows candidates by customer, model, and publication status. Contract decisions are not left to the AI; they are placed as a decision column in a DB view (a virtual table that applies conditions in advance).
- In tests where the AI chooses and uses document search and contract lookup tools, problems occur such as tools not being called, execution errors, and wrong target device errors. The failure cases in the story are fictional examples to explain the design, and they are prevented by checking the call records in code.
- If the work of turning documents into numeric arrays for meaning comparison is done inside the DB, the full set of documents is not sent outside. The AI that generates answers uses an external cloud service.