Everyone in AI is suddenly trying to live on your computer. Today's digest has four distinct threads — and they all converge on that one fact.

Felix Rieseberg builds Claude Cowork at Anthropic. He also helped ship the Slack desktop app and built Electron — the framework that runs VS Code and Discord. His interview today cuts through a lot of noise on where agentic AI is heading.

The central problem he named: a dead-end design choice. Make an AI agent completely safe — and it never does anything real. Give it real power — and you're clicking "approve" on every command. Rieseberg calls the second failure mode "approval fatigue." Neither is acceptable.

Anthropic's answer is a virtual machine — VM for short — an isolated computer running inside your computer. Claude gets its own Linux environment. It can install Python, run scripts, browse the web. You never need to approve individual commands. The container handles the risk.

Here's what's interesting about that. The VM isn't just a safety boundary — it's a capability unlock. Claude can work like a real engineer: installing tools, running loops, making and fixing mistakes — without touching your actual system.

Manus made the same bet this week. The Chinese agentic startup — acquired by Meta — just launched My Computer. A desktop app giving their agent direct access to your local files and terminal. Same thesis: real machine access, real capability.

Perplexity pushed further still. Their Computer agent on Android now controls the local browser directly. Existing cookies preserved. No connectors required. No tool schemas to configure.

Felix made a contrarian point worth sitting with. Silicon Valley is undervaluing local computers. "How come we're all using MacBooks and not iPads?" The machine you work on is where the leverage is. Not a cloud somewhere.

Meanwhile, the economics of running these agents keep collapsing. OpenAI released GPT-5.4 nano this week — twenty cents per million input tokens. Simon Willison ran the math: describing his entire seventy-six-thousand photo collection would cost fifty-two dollars. When inference gets this cheap, the case for always-on local agents becomes hard to argue against.

Simon Willison published a deep breakdown of subagents — and it connects directly to how Claude Cowork actually works.

The problem is context windows. Even at one million tokens, models perform better below two hundred thousand. Stuffing context tanks quality. Subagents — fresh model instances dispatched by a parent agent — keep the root context clean. The parent stays focused. The subagent handles the dirty work.

Willison walks through how Claude Code actually does this. Starting a task on an unfamiliar codebase, Claude dispatches an "Explore" subagent. Fresh context, targeted instructions. It returns a structured summary. The parent agent never burned tokens on that exploration.

What's striking: watching models prompt themselves. Their self-prompting instincts are good. Claude instructing its own subagent to search specific directories, look for specific keywords — coherent task decomposition emerging naturally from the model.

Parallel subagents go further. Dispatch multiple simultaneously on faster, cheaper models — Claude Haiku, for instance — and you get speed and cost benefits at the same time.

This is becoming infrastructure. OpenAI added subagents to Codex this week. Gemini's command-line interface — CLI — supports them. So do Mistral Vibe, OpenCode, and Cursor. The pattern is now standard across the industry.

A paper from this week showed that automatically extracting agent skills from GitHub repos — standardized as SKILL.md files — produced forty percent knowledge-transfer gains. The idea that codified procedural knowledge can travel between agents is getting empirical backing.

Felix Rieseberg's take on skills connects here. Claude Cowork uses "skills" — simple markdown files describing recurring tasks. Not rigid tool schemas. Not complex APIs. Just text. The model figures out the rest. You explain the task the way you'd explain it to a new employee. He's increasingly skeptical of MCPs — model context protocol servers — as the right abstraction. Files and text may outlast them.

Jensen Huang spent two hours at GTC telling what felt like a history lesson. Ben Thompson asked why — straight off the stage. Huang's answer reveals the whole strategic logic.

AI agents will use tools humans built — Excel, SQL databases, design software. None of those tools were designed to run at AI speed. Nvidia's job is to accelerate them. Every tool agents will use needs to get dramatically faster.

The most technically interesting announcement: the Vera CPU — central processing unit. Here's the surprise. Cloud CPUs were optimized for core count — more rentable instances, more revenue per rack. But for inference, where an expensive GPU — graphics processing unit, the chip that runs AI models — is idle waiting on a slow CPU, single-threaded performance is everything. Vera delivers three times the memory bandwidth per core of any CPU ever built.

The Groq acquisition tells a related story. Huang described a fundamental tension: maximize throughput on one end, maximize intelligence per token on the other. These pull against each other. Groq's architecture handles the extreme low-latency end — letting Nvidia disaggregate inference — break the process into specialized stages — all the way down to the token generation layer.

Jensen's line to Thompson: "I would pay for a tier of Claude Code that makes coding ten times faster." He's building this product for himself.

On China, Huang was unusually direct. DeepSeek, Kimi, Qwen are genuinely impressive architectures. Fifty percent of the world's AI researchers come from China. Their open-source work will diffuse globally regardless of export controls. If the American tech stack isn't positioned to receive it, that's a strategic failure — not a victory.

His frustration with AI doomers was equally sharp. Not people genuinely worried about safety. People weaponizing fear to shape policy. "Creating rhetoric that scares people" versus actually warning them. AI's declining public popularity is a real problem. He compared it to early skepticism around electricity — other countries adopted it while America debated.

Andrej Karpathy published — and quietly deleted — one of the most viral AI analyses in recent memory. He scored three hundred forty-two U.S. occupations by AI exposure, zero to ten. The findings: forty-two percent of jobs scored seven or higher. Roughly sixty million workers. About three point seven trillion dollars in annual wages.

High exposure: software developers, financial analysts, writers, editors, graphic designers. Low exposure: construction workers, bartenders, nursing assistants. Roles requiring physical presence that current AI can't replicate at scale.

Karpathy called it "a Saturday morning two-hour vibe-coded project." He deleted the GitHub repo after the debate spiraled.

Felix Rieseberg — building these very tools from inside Anthropic — was candid in a way that landed differently. "We are deeply worried about the impact on the labor market. Especially for junior employees. A lot of work we personally find annoying — that would have been given to a junior entry-level hire." He didn't offer a solution. He named the problem honestly.

Jensen Huang gave a complementary data point. Nvidia's software engineers "haven't generated a line of code in a while." But they're more productive than ever. The architect role is emerging: describe software in specification, not code. Think about structure and systems — not syntax.

And then there's Tim Schilling's quiet warning — worth taking seriously. Writing about LLM use in Django open-source contributions, he said: "If you do not understand the ticket, if you do not understand the solution — your use of LLM is hurting Django as a whole." For reviewers, he added, "it's demoralizing to communicate with a facade of a human."

AI as a vehicle versus AI as a complement. That distinction will matter more, not less.

Here's the thread running through today. The productivity gains are real, verifiable, and accelerating. The displacement isn't hypothetical. And the honest voices — Karpathy, Rieseberg, Huang — are all flagging it while continuing to build. That tension isn't going away. What we do with it is still an open question.