Google has released Gemini 4 Hydrogen, a fast AI model built for agent tasks that automatically handle multi-step work. It finishes jobs 2.5 times faster than the previous lightweight model and scored higher than the larger Gemini 4 Argon on an external overall benchmark.
GitHub's blog introduced agent skills, folders of written instructions for repeatable tasks. Once written, they can be reused by several AI agents. The article explains this with two real examples.
Meta's consumer AI agent Muse can now call U.S. businesses on behalf of users. Only paid subscribers can use the feature. It was first tested in September and is now open to all paid users.
Amazon Cognito now lets services and AI agents acting for a user swap that user's token for a new one. Token exchange means trading one token for a different token meant for another use. Services no longer have to pass along the user's original token as is.
Amazon AWS Machine Learning blog · Nov 13, 06:47 KST
Natera, a leader in cell-free DNA (cfDNA) testing, built an AI system called SWARM with AWS's Amazon Bedrock AgentCore. It automatically diagnoses lab equipment problems by using the know-how of experienced technicians.
Microsoft Research's Focus podcast episode 98 covered AI speech translation, a 31-year-old math problem solved, and efficiency gains for AI coding agents. It also featured a research intern's experience at the Tokyo lab.
OpenAI has released version rust-v0.168.0 of Codex, a lightweight coding agent that runs in the terminal. The introduction does not describe the specific changes.
Google has launched Gemini 4 Hydrogen, a fast AI model for everyday coding and agent tasks. It is meant to replace Gemini 3.8 Flash, and it scored higher than the larger Argon model on several benchmarks.
Google researcher Pietro Miotti said his team of 48 AI agents found a possibly unmapped Maya site in Mexico. The site had already been published in an academic journal last year.
Nectar has launched Creator Agent, an AI agent for brands. It handles finding influencers, contacting them, and managing collaborations and results automatically. The goal is to cut down work that used to take weeks.
Amazon AWS Machine Learning blog · Nov 13, 03:43 KST
Following Part 1 on managing AI spending in Amazon Bedrock, this post asks whether the money was well spent. The key idea is to measure cost per completed task, not the price of a single token. It also covers how to test whether an agent really beats human work.
Amazon AWS Machine Learning blog · Nov 13, 02:59 KST
AWS has upgraded AgentCore runtime, a managed environment for running AI agents. Sessions now start faster, can use more computing resources, and run more efficiently, the company said.
Anthropic released a way for customers to self-host Claude Managed Agents, running the full service and its sandbox on their own cloud accounts or on-premises data centers. Agents, sessions, and session history are stored on that infrastructure, while model requests can go to an inference provider the customer chooses.
xAI's AI agent Grokbot runs on a cloud virtual computer and keeps working even when your laptop is closed. The article also covers how it can reduce repetitive work.
xAI has launched Grok Bot 1.0, an AI agent that runs on a cloud virtual computer around the clock. It can browse the web and prepare drafts on its own, but it can stop when logins expire or approval requests go unanswered.
An article covers xAI's AI agent GrokBot, which keeps running tasks on a cloud virtual computer even when the laptop is closed. It looks at how much repetitive work it can take off people's hands.
United Airlines will restrict AI agents, programs that act on a user's behalf, from booking flights and managing reservations. The airline says handing account logins to AI assistants can create security problems.
TikTok's parent company ByteDance is reportedly developing its own AI agent, an assistant that does tasks for users instead of only answering questions. Bloomberg says it could launch within weeks.