Agents' Last Exam is a large-scale benchmark built by Berkeley RDI and over 300 industry experts, covering 55 sub-industries and 1,500+ tasks to evaluate AI agents on long-horizon, economically valuable, verifiable real-world workflows. It measures agent performance in professional software like Adobe After Effects, Siemens NX, and Unreal Engine across animation, engineering, game development, manufacturing, architecture, and neuroscience.

Cortex introduces the DRIVE framework to measure engineering organizational health in the AI era. It assesses effectiveness across five pillars—Delivery, Reliability, Initiatives, Vigilance, and Efficiency—and uses recurring Operational Excellence reviews to turn measurements into action.

OpenAI has released its GPT-5.6 family of models, including three variants: Sol, Terra, and Luna. The new models compete with Anthropic's Fable 5 on benchmarks but at lower cost, with improvements in coding, knowledge work, and safety.

A new economic analysis from theCUBE Research found that migrating from a fragmented, on-premises data estate to Oracle Corp.’s Autonomous AI Database on Multicloud can deliver meaningful modernization value — and substantially larger returns when organizations use the platform to accelerate AI projects.

The updates amount to a significant expansion of the platform.

Samaya Research introduces FrontierFinance, the largest open benchmark for investor workflows.

Frame.js is an open-source, browser-embedded runtime for creating interactive apps, charts, and sketches. Use AI or code, then share via self-contained URLs instantly. No build, server, or account required.

Google Cloud announces the public preview of Cloud Run sandboxes, a native, secure, and ultra-fast runtime environment for executing untrusted code and agent workloads, starting in milliseconds. It supports use cases like LLM code interpreters, headless browsers, and user-submitted code execution, with zero-trust security through credential isolation, default-deny egress, and read-only filesystem overlay.

Autopilot-Jobhunt is an AI tool that scans job postings, matches them to a user's resume, and notifies via Telegram. It can also generate tailored resumes and cover letters, but never auto-applies. It uses free AI models and prioritizes privacy.

Kastra is a platform that provides runtime authorization for AI coding tools like Claude, Cursor, Codex, and OpenClaw, aiming to enhance security and control.

The model release is the first since SpaceX went public in June and will help SpaceXAI compete with other frontier model providers, particularly in coding.

ChatGPT Work is an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.

AI agents are evolving from one-time assistants to persistent workers that can repeat tasks, monitor changes, and run checks until a goal is met. This article explores the concept of agent loops, focusing on Claude Code's /loop command and OpenAI Codex automations, covering architecture, loop types, and hands-on examples.

Microsoft is accelerating AI-driven vulnerability detection in Windows. Its new pipeline found 16 vulnerabilities in May, with more fixes expected per update. Enterprise admins need to adapt to faster patching cycles with tools like KIR and hotpatch.

About two weeks after a limited preview restricted to government-approved organizations, OpenAI has received approval for public rollout of GPT-5.6. Simultaneously, the company unveiled ChatGPT Work, an AI agent combining ChatGPT and Codex for non-technical users. Powered by the GPT-5.6 suite (Sol, Terra, Luna), it integrates with various apps to create documents, spreadsheets, etc. The desktop app is available globally, with mobile and web access rolling out gradually. OpenAI aims to lead in the competitive AI agent space.

Windows 11 updates could soon include fixes for more security issues at once. Microsoft said in a blog post on Thursday that it's now using AI to “identify potential issues earlier,” which means “customers will see a higher volume of security updates included in each security release.”

A new Dartmouth study finds that AI-drafted responses to patient messages often contain errors and extraneous details, forcing physicians to spend more time editing than if they had written the messages from scratch.

As AI agents generate performant GPU kernels, building trust becomes the bottleneck. Gimlet Labs presents an early research system that uses formal verification to catch bugs missed by numerical tests, demonstrated with a missing clamp in attention kernels.

As agentic AI accelerates enterprise transformation, data sovereignty is crystallizing from a compliance checkbox into a foundational strategic imperative — one that determines not just where data lives, but who captures the economic value it generates. The debate is particularly acute in Europe, but companies globally are reassessing control over hyperscalers, model providers, and closed-weight systems.

As AI models grow larger and costlier, users report declining quality and rising bills. This article coins the term 'model bloat' to describe the accumulation of unnecessary size and complexity without proportional gains, and argues that naming the pattern helps address it.

Aurora 1.5 adds 22 more variables, hourly temporal resolution, and probabilistic ensemble forecasting to the Aurora foundation model, making it more useful for real-world weather, climate, and energy applications. Released as open source, it enables researchers and developers to use, evaluate, and build on the model.

In this post, we show where MCP tool design goes wrong and how to fix it with practical context engineering approaches.

This post walks through five new capabilities for Amazon SageMaker HyperPod inference: multi-tier data capture for auditing and model improvement, direct deployment from Hugging Face Hub, local NVMe model loading for faster cold starts, automated Route 53 DNS for custom domains, and pod-level IAM through custom service accounts. These enhancements provide faster, more observable, and more flexible inference infrastructure for enterprise generative AI workloads.

Meta introduces Muse Spark 1.1, the first Spark model with an API, featuring improvements in agentic tool calling and computer use. The evaluation report reveals interesting 'attractor states' in self-conversation. Developer Simon Willison created a plugin for CLI access.

Simon Willison released llm-meta-ai 0.1, an LLM plugin for the Meta AI API, enabling prompts against the new muse-spark-1.1 model.

Simon Willison releases llm 0.31.1, fixing a JSON error from some providers when an OpenAI Chat Completion tool call has empty arguments.

Learn how to build a venture capital research agent that produces investment memos in 90 seconds with cited sources, using the Perplexity Agent API, LangGraph, and LangSmith. The agent runs parallel research nodes for team, financials, product, and market, then synthesizes a memo with seven sections, including a thesis and recommendation. Every claim is traced to primary sources, making the output auditable. The article also compares three search providers and offers takeaways for building similar agents.

A new report by Pangram Labs reveals that over a quarter of long-form social media posts are fully AI-generated, with LinkedIn being the most affected platform, where over 40% of long posts are AI-generated. The study is based on over 1 million posts analyzed via an opt-in Chrome extension.

An open-source AI framework that uses Claude Code to automate job applications: evaluating postings, tailoring CVs, writing cover letters, and preparing for interviews.

Questioneverything.ai is a platform that uses composable knowledge blocks to help users turn curiosity into knowledge, enabling structured exploration of any question.