Summary
The morning feed was dominated by one cluster: NVIDIA and Microsoft using GTC Taipei and Microsoft Build to push AI agents onto the Windows PC, eight NVIDIA posts spanning the OpenShell secure agent runtime, a new "CPU for agents" called Vera, the Nemotron 3 Ultra open model, DGX Station and RTX Spark hardware, and Claude models running natively on GB300 Blackwell in Azure. The single most informative artifact is NVIDIA's OpenShell diagram, which shows a local-to-cloud query router sitting between a prompt and a shelf of local models. A second, related cluster is the move to run agents on-device: Nous Research shipped Hermes Desktop in public preview and Scoble framed "Hermes vs OpenClaw" as the new Mac vs Windows. A third cluster is coding-agent tooling: Anthropic's ClaudeDevs rolled out a Claude Platform CLI ("ant") and a revamped /fork, Cursor published lessons from a year of building cloud agents, and Kilo Code argued the real limit on agentic engineering is human review bandwidth, not agent count. Standout non-cluster items: Sholto Douglas (Anthropic) backing NewLimit's $435M longevity Series C, and a Kilo bug-finding bake-off where Opus 4.8 led at 10 of 15 planted bugs. The rest of the feed was heavy political and promotional noise (@brivael, @spencerpratt, @WHFraudTF) with no AI research substance.
Posts
NVIDIA and Microsoft reinvent the PC around agents (cluster of 8, @nvidia, blog). NVIDIA's GTC Taipei / Microsoft Build posts describe a full agentic-AI stack from Windows device to Azure cloud to local. The headline for this wiki is OpenShell, a secure-by-design agent runtime: the attached diagram, titled "NVIDIA and Microsoft Reinvent PC," shows a prompt feeding an OpenShell box that runs a Context to Observe to Reason to Act loop, with Claude and a CLI on one side and a shelf of local models on the other (DeepSeek, Gemma, GLM, GPT-OSS, Kimi, MiniMax, Nemotron, Qwen) sitting on top of Windows. NVIDIA describes OpenShell as providing governance, policy enforcement, and "smart local-to-cloud query routing," which is a production deployment of the routing-as-policy idea. Adjacent posts: NVIDIA Vera, billed as "the CPU for agents" with 80% faster agentic task completion than x86; Nemotron 3 Ultra, a smaller open model with 5x faster inference and up to 30% lower cost for long-running agents; RTX Spark Windows PCs and DGX Station for local data-center-class compute; and Claude models running natively on GB300 Blackwell Ultra in Azure. See LLM routing and the 06-03 digest.
Hermes Desktop ships, "the new Mac vs Windows" (@ns123abc reposting @NousResearch; @Scobleizer). Nous Research released Hermes Desktop in public preview, a native on-machine version of its Hermes agent first demoed in Jensen's GTC keynote. Scoble framed the moment as "Hermes vs OpenClaw is the new Mac vs Windows," casting the two agent platforms as the rival desktop ecosystems of the agentic era. This is the on-device-agent counterpart to NVIDIA's OpenShell push above: the same week, the local PC became the contested surface for agents.
Claude Platform gets a CLI and a smarter /fork (cluster of 8, @ClaudeDevs, docs). Anthropic shipped "ant," a CLI that makes every Claude Platform API endpoint runnable from the terminal: Messages API, Managed Agents (create, send an event, pull the full trace of every tool call and decision), and files, with typed flags or piped YAML and a TUI output mode. It is designed to be driven by coding agents through a built-in "claude-api" skill, and supports version-controlling an agent as YAML synced from CI. Separately, /fork was updated to run a background agent with your exact context (system prompt, tools, history, model) and prompt cache, returning the result to your session, while the old behavior moved to /branch.
Cursor's lessons from a year of cloud agents (@cursor_ai, blog). Cursor argues a good cloud agent is much more than a local agent moved to a server: it needs a durable execution platform, a powerful harness, and realistic development environments. Cloud agents now run on dedicated VMs with their own dependencies and network access, can work in parallel and unattended, and the engineering increasingly looks like building an operating layer around the agent rather than porting one. A concrete companion to NVIDIA and Anthropic's runtime announcements on the cloud side of the same problem.
Kilo Code: the real limit is human review, not agent count (cluster of 2, @kilocode, blog). Kilo's lesson from 40 trillion tokens and 3M downloads: "if a human cannot review the output in one sitting, the task was too big," so the useful version of agentic engineering is 2 or 3 agents you actually steer, not 100 in 100 tabs. A second post warned that coding agents get ripped out when vendors pivot (Roo Code archived, Google's Antigravity bait-and-switched into a chatbot overnight) and argued for workflows that survive a tool swap. Practitioner counterweight to the agent-everywhere marketing.
Kilo bug-finding bake-off (@kilocode, via @coldopn). On 15 planted bugs with one prompt, Claude Opus 4.8 caught 10, Grok Build 0.1 and Claude Sonnet 4.6 each caught 9, GPT-5.5 caught 8, and Gemini 3.1 Pro caught 2, with Grok Build 0.1 finding the hardest three-file state-mutation chain at about a third of the cost per catch of its nearest rival.
Sholto Douglas backs NewLimit's $435M longevity round (@_sholtodouglas reposting @newlimit). Anthropic's Sholto Douglas called NewLimit "the most exciting company in biology" and doubled down on its $435M Series C, led by Founders Fund, to bring cell-age reprogramming (reprogramming biological cell age to extend healthy years) into human trials. Notable as an AI-frontier researcher putting weight behind a longevity-bio bet.
Genomi: a local DNA-expert agent harness (@Scobleizer reposting Hongjian Zhou). An open-source agent harness that turns a coding agent into a personal DNA expert, pitched on three claims: general AI can sound right while being wrong about genomics, static DNA reports cannot keep up with new science, and DNA data should stay on the local device rather than be uploaded. Fits the on-device-agent theme of the morning.
Tencent WorkBuddy and screen-watching work agents (@Scobleizer). Scoble flagged Chinese "do a whole week of work" agents (Tencent WorkBuddy) and a Texas screen-watching assistant (Timeglass), noting the privacy and security unease but conceding they help a lot, and predicting mass adoption "after Microsoft just does the same."
AI IPO race newsletter (@Scobleizer, unaligned.io). Scoble and Irena Cronin's free issue argues AI companies have entered a capital-intensive stage where Anthropic, OpenAI, and SpaceX need huge sums for data centers and compute, and the open question is whether public investors will believe the valuations. Industry-watching, no research substance.
Promotional and political noise (Skip). @brivael (19 posts, almost all French political commentary plus an Argil AI-video-app promo), @spencerpratt (4 LA campaign ads), @WHFraudTF (anti-fraud politics), @imagine (Grok Imagine promo), @tinygrad ("the spice must flow"), and @lexfridman (Jensen-in-Taiwan travel color) carried no AI research substance this slot.