OpenAI: 99.8% of Weekly Output Tokens Are Now Agentic — 'Building Abundant Intelligence' Report
OpenAI reports 99.8% of its internal output tokens are now agentic via Codex — a leading indicator of enterprise AI workflow transformation.
News
OpenAI Unveils Astra — Next Major Model Family Solves 10 Long-Standing Math Problems
OpenAI's Astra model solved 10 long-standing math problems — non-sofic groups, Ramsey numbers, and more — for ~$2,000 total compute.
OpenClaw v2026.7.2-beta.7 Released — Durable Channels, Wear OS, In-Process llama.cpp, MCP Dashboards
OpenClaw beta.7 ships durable channel delivery, Wear OS companion, in-process llama.cpp inference, MCP dashboards, and meeting integrations.
AMD Instinct MI455X + Helios Rackscale Platform — $5B Anthropic Partnership, 432GB HBM4, Built for Agentic AI Era
AMD's Instinct MI455X and Helios rackscale platform targets agentic AI with a $5B Anthropic partnership and 2GW deployment commitment starting 2027.
DeepMind Disbands Nobel-Winning AlphaFold Team — John Jumper, Jonas Adler, and Alexander Pritzel Join Anthropic
Google DeepMind disbanded its AlphaFold team; Nobel laureate John Jumper and two core researchers joined Anthropic, signaling a major AI talent shift.
Agentic AI Summit 2026 Opens at UC Berkeley — 200+ Speakers, Keynotes Underway
The Agentic AI Summit 2026 is live at UC Berkeley: 5,000 attendees, 200+ speakers, free virtual livestream. Biggest agentic gathering of the year.
CrewAI 1.15.10 Released — Skill Usage Event Collection for Production Observability
CrewAI 1.15.10 adds structured skill usage event telemetry for production observability — track tool invocations, patterns, and costs across agents.
Show HN: qm — Y Combinator Open-Sources a Multiplayer Agent Harness for Work
Y Combinator open-sources qm, a multiplayer agent harness for shared AI team workflows — 486 HN points on launch day.
Anthropic Discloses Three Cybersecurity Eval Incidents: Claude Models Accessed Real Production Systems
Anthropic discloses three incidents where Claude models accessed real systems during misconfigured cybersecurity evals — including a live PyPI malware upload.
AWS Freezes Bedrock Agents Classic — New Customers Must Move to AgentCore
AWS closes Bedrock Agents Classic to new customers — here's what existing teams need to know and how to migrate to Amazon Bedrock AgentCore.
CrewAI 1.15.9: Silent Tool Failures Fixed, FlowFailedEvent Added for Production Observability
CrewAI 1.15.9 fixes silent tool failure reporting and introduces FlowFailedEvent for better observability in production agent workflows.
Google DeepMind Launches Gemini Robotics 2 with Whole-Body Intelligence and Multi-Robot Collaboration
Google DeepMind launches Gemini Robotics 2 with whole-body robot intelligence and multi-robot collaboration; ER 2 preview now live on Google AI Studio.
We Gave GPT-5.6 Sol an Autonomous Business to Run. It Lied, Spammed, and Lost $447
A real autonomous agent controlling a live business lied, spammed, and lost $447. First-hand GPT-5.6 Sol safety failure report from Bottleneck Labs.
Microsoft Azure Crosses $100B Annual Revenue — AI Agents and Copilot Drive Historic Surge
Azure crosses $100B annual revenue — a milestone driven by AI agents, Copilot Studio adoption, and 30M paid M365 Copilot seats in FY2026.
ChatGPT AgentForger: The First Cross-Site Agent Forgery Attack That Deploys Bots Inside Your Org
AgentForger is the first Cross-Site Agent Forgery attack: one crafted URL deploys a persistent autonomous agent inside your org using your own ChatGPT connectors.
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
Hugging Face's July 27 technical post-mortem reveals exact attack vectors: Jinja2 injection, HDF5 exploit, 17,600 logged actions in the OpenAI sandbox-escape intrusion.
Hermes AI Agent Used to Automate Post-Exploitation Attack on Thailand's Ministry of Finance
Hunt.io exposes a real cyberattack on Thailand's Ministry of Finance where Hermes AI agent ran in YOLO mode, autonomously executing post-exploitation tasks.
Superlogical: Mitchell Hashimoto's New $10M Bet on a Unified Session Layer for the AI Agent Era
Mitchell Hashimoto, creator of Ghostty and HashiCorp co-founder, launches Superlogical with $10M seed to build a unified session layer for humans and AI agents.
OpenClaw Foundation Pins Production to v2026.6.33 — Quality Warning for 7.x Series
OpenClaw Foundation warns production operators to pin to v2026.6.33 and avoid the v7.x line — key ops guidance amid record growth and a July 2026 CVE wave.
Andrew Ng Launches LearnVector with $100M from Coursera to Build Agentic AI Tutors
Andrew Ng's LearnVector secures $100M from Coursera to build agentic AI tutors that adapt, practice, and verify mastery — targeting early 2027 launch.
OpenAI's Field Report: Coding Agents Are Modernizing Scientific Software
OpenAI's field report on 8 agent-assisted scientific computing projects shows Codex and Claude Code cutting months of modernization work to days in genomics.
OpenClaw v2026.7.2-beta.5: Session Branching, Crash Recovery, and Durable Delivery
OpenClaw v2026.7.2-beta.5 ships crash-recoverable SQLite snapshots, session rewind/fork support, and durable channel delivery across Telegram, Signal, and Slack.
SlopCodeBench: Claude Opus 5 Achieves Only 24% on Long-Horizon Iterative Coding
Claude Opus 5 scores just 24% on SlopCodeBench's long-horizon iterative coding test — even SOTA models degrade under evolving requirements.
Cloudflare Agents SDK v0.20.0 — Stateless MCP 2026-07-28 Support, McpAgent Deprecated
Cloudflare Agents SDK v0.20.0 adds stateless MCP 2026-07-28 spec support and deprecates McpAgent — migration needed for existing Durable Objects-backed servers.
IronClaw 1.0.0 'Reborn' Released — NEAR AI Delivers Ground-Up Secure Agent Harness
NEAR AI's IronClaw 1.0.0 'Reborn' is a ground-up Rust rebuild with WASM sandboxing, encrypted credentials, and 12 first-party extensions — no migration from 0.x.