This page is only in Korean for now. Your browser can translate it, or go to the English home.
Technologies
목표를 받고 도구를 써서 여러 단계를 수행하는 AI예요.
대화에서 실제 업무 실행으로 옮겨가는 흐름과, 사람이 확인해야 할 경계를 함께 봐요.
Up to 40 recent matching stories, newest first. Newly collected stories appear here automatically.
Gathering stories…
Search earlier stories →Names are matched in the titles and summaries of public stories collected by LastAI. Some stories may be missed or refer to a namesake. Coverage is not exhaustive. Profiles updated .
Some data could not be loaded. These are the stories available so far; try again shortly.
Korail has blocked the AI agent Pickle, which plans and carries out tasks on its own. Pickle can no longer book or operate train tickets through Korail.
This article explains how Java teams can use Amazon Bedrock and Spring AI to build AI agents that plan steps, call tools, and remember conversation context. It also notes that moving such agents from prototype to production is difficult.
United Airlines will not let personal AI assistants (programs that act on a user's behalf) book its flights. The assistants can still show flight options, but they cannot complete a purchase.
AWS's machine learning blog introduced UAEF, an evaluation system for AI agents. It scores agents built on different frameworks with the same metrics and compares agent versions. Organizations in telecom, healthcare, and financial services used it before releasing their agents.
Databricks has released an updated version of its AI agent Genie. It can carry out long, multi-step tasks without human supervision, using an ontology (a structured map of how data relates) built from customer data.
AI agents, which handle multi-step tasks on their own, are still considered too unreliable to run without close supervision. Startup Plurai says it will use the same kind of simulation that helped bring self-driving cars to roads to tackle this problem.
UC Berkeley and NVIDIA released open-source tools that let teams of AI coding agents write, simulate, and refine robot programs. Setting up industrial robots usually takes weeks to months, and the researchers say this approach could cut that time.
ElevenLabs launched a new partner program to help companies adopt AI agents, which are programs that do tasks for people, faster. Partner firms get sales incentives, new certifications, and technical support.
Microsoft researchers found a backdoor technique called Lazarus that reprograms a few dormant neurons in an AI model. The model behaves normally until a hidden trigger appears, then runs malicious commands. The risk is highest for AI agents that have access to tools or terminals.
OpenAI has launched a beta feature in ChatGPT that shows business information, competing with Google's business listing service. Companies can manage their information and products inside ChatGPT.
eBay seems to have blocked outside AI agents, including Meta's Muse, from browsing its site. A reader's Muse agent had been compiling daily watch listings for weeks, but this weekend it could no longer get eBay details.
Workday built an internal platform for creating and running AI agents, which are programs that carry out work on their own, for HR, finance, and IT. The team used LangGraph to design each agent's workflow and LangSmith to trace and evaluate how the agents behaved.
Meta has released its Muse AI assistant on Apple Watch and Google's Wear OS smartwatches. Users can now talk to Muse from their wrist by voice call, text, or dictation.
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.
This video covers how to use the KV cache (a space where a model stores calculation results during a conversation) as a security signal for AI agents. The speaker is Ari Azarafrooz, founder of Intrinsec AI. The video was posted on the AI Engineer channel on November 9, 2026.
Stanford researcher Maty Bohacek examines how AI agents (AI programs that act on their own toward goals) handle procedures that verify someone is human through phone or video calls. The scenarios include opening accounts, approving money transfers, and recovering access.
Mislav Ivanda of Daytona shares what he learned from adding sandboxes (isolated execution environments) to major agent frameworks. Frameworks that disagree on almost everything have reached the same conclusion.
This talk covers how AI agents copy accessibility problems found in the code they learned from. Accessibility means making products usable for people with disabilities without barriers. In an experiment, the speaker says an agent given accessibility rules finished the same task faster and at lower cost.
Ashwin Paranjape of Samaya AI explains why forecasting company earnings is a better test for AI finance agents than predicting stock prices. He says the test ran on more than 450 earnings announcements in an environment that blocked access to future data.
David Ondrej's channel has a video that explains what Linux is, a type of basic operating system that runs computers. It also covers why AI agents, programs that carry out tasks on their own instead of by people, tend to prefer Linux.
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.
OpenAI has released a new test called BioBasicsBench. It uses 100 tasks to measure how well AI agents (AI programs that handle several steps on their own) can read raw experimental data and carry out basic lab work.
NVIDIA has built a system in which multiple AI agents work together to help run its own AI data centers. The agents find problems, connect the causes, and show recommendations to the responsible engineer through chat. The goal is to reduce the effort of checking multiple dashboards one by one.
NVIDIA has introduced the FOX Blueprint, a reference design for building AI agents that manage factories. Multiple specialized agents work together to inspect defects, assign repairs, and investigate causes. It can also be installed so that it runs only inside a factory's internal network.
The AI agent feature of Amazon WorkSpaces now supports Linux and Windows 11 cloud desktops, in addition to the existing Windows Server option. Each agent can use the operating system that best fits its task. Agents can run without changing existing work environments.
A new method has been introduced to make powerful AI follow rules. Instead of looking inside AI to understand it, this approach designs systems so that following the rules becomes the more profitable choice. The article summarizes the ideas of the researchers who proposed this approach.
