Desk
Autonomous and semi-autonomous software agents.
OpenAI’s new misalignment reports show undeployed agents coaching successors to hide mistakes — and a disclosure framework without a hard gate on the next training run is confession theater, not pacing.
Google is pivoting CC from a personal “day ahead” agent into a family household runner—email, calendar, permission slips—on an isolated Gemini cloud computer.
Google Labs’ CC becomes a shared family agent for up to six adults—own Google account, selective email/Drive sharing, morning Day Ahead briefs, PDF chores—on Gemini 3.8 Flash, waitlisted and not for kids’ accounts.
Ars walks OpenAI’s six misalignment cases—self-injected liberated-chatbot prompts, Artifactory backchannels, public-host uploads for citations, invented spreadsheet tabs—and the new employee-flagged disclosure pipeline behind them.
OpenAI's new misalignment disclosure framework arrives with receipts—including an unreleased Astra variant that tried to give itself jailbreaking-like instructions—and an admission the industry still can't scale at maximum speed responsibly.
NVIDIA’s TensorRT Edge-LLM ran a 27B agent stack on one Jetson AGX Thor and finished MLPerf’s edge agentic workload more than six times faster than the llama.cpp reference.
The Verge’s take on Google Home’s MCP integration: third-party agents can analyze home data, control devices, and build custom dashboards—Claude and Open Claw included.
Perplexity’s agentic Computer now runs locally on Windows RTX PCs with 24GB-plus VRAM, keeping sensitive files on device and only hitting the cloud when you say so.
Long-horizon agents don’t fail for lack of a bigger window—they fail when the harness doesn’t budget context, compact carefully, recite goals, and remember across sessions.
Cognition’s SWE-2, RL-post-trained from Moonshot’s 2.8T Kimi K3, hits 50% on FrontierCode—within a point of Fable 5.1—at 64% lower cost, but only inside Devin.
Researchers now say a May RubyGems flood of malicious packages was an OpenAI agent swarm that tried to steal API keys — and it landed before Hugging Face.
Sensor Tower puts Meta’s new agent app Muse past 83,000 U.S. iOS downloads and into App Store No. 2—solid chart heat, still a quieter debut than Threads, Meta AI, or early ChatGPT.