Desk
Autonomous and semi-autonomous software agents.
NVIDIA's walkthrough of agents helping design AI data centers is most useful for what it admits: the agent drafts, and engineers still have to check every dimension.
Mistral's Voxtral 3 is a cheap, credible full voice stack with a European hosting story, though realtime accuracy still trails Google and no open weights shipped.
llm-d 1.0 is unglamorous plumbing, and that is the point: agents are making inference a routing problem, and an open, multi-vendor answer just reached stable.
A hobbyist's openly declared AI agent got a mailing list of internet scholars to panic in under an hour, and the reaction says more about us than about the bot.
Terminal X is trying to turn senior investors' judgment into versioned, auditable agents, and the most interesting part isn't the model but the grading loop around it.
A breach at OpenCode's hosted console drained stored AI keys overnight, and its defenders found their own AI assistants refusing to help mid-fight. Both halves of that story should worry us.
Meta found its all-knowing agent hallucinated in production, so it built a dumb router and fourteen narrow specialists. That is the most useful agent lesson we have read this month.
Meta lets its adults-only Muse agent sext with users while the human performers who make adult content keep fighting restrictions on Instagram. That gap is the story.
MarkTechPost’s take on Nemotron 3.5 Super VL: 120B open multimodal MoE, 12B active, 1M context, NVFP4 weights aimed at agents.
Meta’s Forum app was a testbed for human-on-the-loop agents — success measured as verified production changes, not lines of code.
The Verge on Muse: you may have clicked yes — your friends never did — and security researchers say the agent builds hourly dossiers on everyone in your life.
CopilotKit’s CLiF turns agent traces and user actions into skills, evals, and fine-tuning data — closing the loop agents keep failing in production.