This part covers how to expand the work agent built in Part 1 into an environment where multiple agents and tools are used. A gateway at the entrance receives all requests and first filters them using policies. A second agent running on another platform is connected in the same way.
Amazon AWS Machine Learning blog · Nov 7, 02:09 KST
This article explains how administrators can control who can access what when AI agents reach business systems on behalf of users. It uses Amazon's AgentCore together with Okta's Cross App Access (XAA) feature. This is the first part of the series.
Google has added a voice mode to its Antigravity AI coding tool. Users can give tasks by speaking and change direction while the agent is working. The voice mode runs on Gemini Live, Google's real-time voice AI model.
Amazon AWS Machine Learning blog · Nov 7, 01:55 KST
Greenlight Financial built Luna, a security incident handling system, on Amazon Bedrock AgentCore, a platform for AI agents (programs that make their own decisions and carry out tasks). Luna handles more than 90% of up to 4,000 security and risk incidents per month without human involvement. The article explains the system's structure and how it operates.
This article covers the problem that coding agents (AI tools that read and fix code for you) can read every file in a repository but often do not understand how the app actually runs. It explains that information such as data connections, deployment locations, and the cause of last night's failure lives on platforms outside the code.
Insilico Medicine has released Longevity Grok, an open-source system in which multiple AI agents organize health records, suggest interventions, and score the results. It was built with Harvard Medical School researcher David Sinclair. The system is restricted to research and educational use only.
This guide explains AI evals, the methods used to find out whether AI agents actually work well. It explains how they differ from regular software testing and introduces an approach that evaluates agents with code checks, LLM judges (where another AI model grades the answers), and human review.
LangChain has added browser features to its managed Deep Agents. Agents can now handle web tasks that have no API by looking at the screen and clicking, like a person would. Users can also set how many browsers are opened and how long they stay open.
As AI agents take on bigger goals, they need a way to run several long tasks at once while still talking with the user. LangChain has introduced this into Deep Agents by splitting the system into a 'coordinator' that handles the conversation and 'asynchronous subagents' that work in the background.
The LangChain team chose a browser tool that lets AI agents operate websites directly, then improved it by reviewing real records of agent work. After comparing several tools, they picked Stagehand and checked the effect of each change with tests.
AI agents can only hold a limited amount of information at once, so they must find and bring in the information they need. LangChain explains how this retrieval becomes harder and introduces ways to reduce the difficulty. It uses a competitor analysis agent as an example.
A voice AI agent answers phone calls and speaks through several connected layers of technology. This article explains how those layers work together. It uses ElevenLabs' service ElevenAgents as an example.
Cogent Security, a US startup, released VR-1.1, an AI model built for security work. It scored higher than Anthropic and OpenAI models on one specific security benchmark. However, its cost efficiency was lower, and this result alone is not enough to call it overall better.
Yardlet is a Mac app that lets coding agents handle web logins and pass messages between tasks on your behalf. It is a native app built with Swift and released under the MIT license. The developer built it for daily use while switching between multiple agents and accounts, and maintains it alone.
Meta has added real-time voice conversation and video calling to Muse, its personal AI agent. On screen, the AI avatar changes its facial expressions and gestures to match the conversation. It is first available to users in the United States.
Kakao has released its in-house decision-support model, Kanana-2-v-decide, on Hugging Face. The company says each request takes an average of 0.04 seconds or less to process. The model produces results in fixed formats such as probabilities and scores rather than sentences, so services can use them right away.
OpenAI's coding tool Codex has released version 0.167.0. Users can switch between multiple backends (the connected AI service servers) from the terminal, and the agent command center shows results that have not yet been checked. Image preview and a Google Cloud Vertex AI setup wizard have also been added.
Meta has started giving all US users access to Muse Voice, a feature that lets its personal AI agent Muse talk in real time, like a phone or video call. The feature was first shown at Meta's developer event in September.
Kakao has released 'Kanana-2-v-Decide', a multimodal model that helps AI agents make decisions, on the open-source platform Hugging Face. The model looks at both text and images and reports results in a set format.