Google Cloud has added two new systems to its 8th-generation TPU lineup. One adds hardware-based confidential computing to protect enterprise data, while the other expands memory so AI agents can handle large volumes of material at once.
This talk covers how to make AI agents (programs that split up and handle tasks on their own) reliably help with money matters. Presenters use wedding costs as an example and build a financial analysis agent live. They point out that a trustworthy answer is harder to get than a fast one.
When coding agents send a flood of requests at once, a self-hosted LLM server may accept them all and end up pushing out the ones that matter most. This video covers admission control, which decides what to reject before requests reach the GPU (graphics processing unit).
Online payments company Stripe says it built Kai, an AI assistant used by all of its employees, in one week using a tool called Deep Agents. It is designed so that people who are not developers can ask it to analyze data, write documents, and organize ideas.
Chinese AI company DeepSeek's autonomous agent framework, Harness, is growing its user base quickly. Harness is foundational software that helps AI work with files and run programs directly. The company plans to use this momentum to begin monetization through paid services.
AI developer YouTuber David Ondrej has released a video explaining how he uses agent skills (collections of instructions that teach AI how to perform specific tasks). He argues that people should build their own skills rather than copying others'.
Cisco has announced that it will add Anthropic's Claude Mythos model to AI Defense, its security tool for checking AI apps and agents. It is used for red-team testing (deliberate attacks to find weaknesses) and works only within approved targets. The feature will soon be available to some enterprise customers.
Meta has released Muse Ember, its smallest and most efficient AI model. It is built to handle the simple, everyday tasks that AI agents (AI programs that automatically carry out multiple jobs on behalf of people) perform repeatedly. It can process up to 1 million input tokens (small units of text) at once.
Amazon AWS Machine Learning blog · Nov 10, 04:05 KST
AWS has introduced a seven-layer reference architecture for running multiple AI agents (AI programs that handle tasks on their own) at real-service scale. It notes that prototypes built for testing often cannot handle large volumes of requests or pass security checks. The architecture aims to provide a shared foundation that all agents can use.
Amazon AWS Machine Learning blog · Nov 10, 03:55 KST
CrossnoKaye built ATLAS, a platform that controls factory equipment using real-time data and AI, and AURA, an AI system that runs on top of it. Attaching AI to industrial sites, where equipment and systems from many companies are scattered, is not easy. This article describes how the company designed this structure on AWS.
LangSmith, a tool for checking AI apps, now shows GitHub pull request (code change proposal) results and spending on its coding agent overview. Users can set an agent to Coding type and connect GitHub accounts.
Meta has released Muse Ember, its smallest AI model. It is designed for tasks that AI agents repeatedly perform, such as tool calls, coding subtasks, and document processing. It is priced at $0.10 per million input tokens.
Ryan Carson, founder of Untangle, shows how his team builds software using AI agents (AI programs that carry out tasks on their own) in the cloud. One management agent leads up to 20 development agents.
Meta Superintelligence Labs has released Muse Ember, its lightest and most cost-efficient AI model. It is designed for tasks that AI agents (AI programs that handle multi-step work on their own) repeat throughout the day, such as tool calls and code subtasks.
Several teams are grouping multiple AI agents under one coordinating role to handle tasks such as risk assessment and customer analysis. The article asks how this kind of setup can uphold zero trust (a security approach that verifies every access request each time), which holds that no one is trusted by default and everything is always checked.
LangChain introduces eval-driven development (EDD), in which developers set evaluation criteria (evals) first and build the AI agent to meet them. Improving agents by feel becomes harder to judge over time, which is why this approach is useful. The article explains the process step by step.
Meta has acquired Conductor, an app that runs multiple AI coding agents at the same time. The acquisition price and detailed reasons were not disclosed. The Conductor team will join Meta Superintelligence Labs to build its next product.