Claude Code, Cowork and Codex #6: Claude Code Auto Mode and Full Cowork Computer Use
TL;DR
Anthropic’s biggest quality-of-life win is Claude Code Auto Mode — instead of approving every file write and bash command or going full “skip permissions,” a classifier now auto-allows safe actions and blocks risky ones, and Adam Wolf says it has become “core to how I work with Claude.”
Claude Co-work can now fully use your computer — keyboard, mouse, screen, everything — and paired with Dispatch or channels like Telegram and Discord, that means you can effectively operate your desktop from your phone, which the creator frames as both extremely useful and extremely dangerous.
The video’s core warning is simple: don’t YOLO full computer use on a machine that matters — if Claude has access to your default Chrome profile, email, or sensitive accounts, prompt injection, remote compromise, or phone takeover could turn convenience into “an extremely bad time.”
Anthropic is shipping at a ridiculous pace in 2026 — the recap lists launches from Claude Co-work and Opus/Sonnet 4.6 to a 1M context window, marketplace, Dispatch, projects, computer use, and Auto Mode, ending with the blunt summary: “Anthropic is cooking.”
Claude Code Review looks like a serious product, not a toy — Anthropic says large PRs over 1,000 lines average 7.5 issues found, with less than 1% marked incorrect, and reviews typically cost $15–$25, which the creator says is often obviously worth it if engineer time is the real bottleneck.
Agentic coding is crossing from novelty into real workflow infrastructure — examples range from Gary Tan’s GStack setup and Dean Ball’s nearly 10-hour Codex sessions to custom tools for transcription, research, podcast editing, and even a piano-learning app, with the new heuristic shifting from “if you do it 3 times, code it” to basically “build it the first time.”
The Breakdown
Auto Mode Is the Missing Middle Between Babysitting and Chaos
The video opens on Anthropic’s new Claude Code Auto Mode, which sits between manually approving every action and dangerously skipping permissions altogether. A classifier checks each tool call, lets safe actions through, blocks risky ones, and only escalates when needed; Adam Wolf calls it a “total gamechanger,” while the creator jokes that yes, some people will still mindlessly click “yes,” and those people are fools.
Full Computer Use Arrives — and the Creator Immediately Sounds the Alarm
Next comes Claude Co-work’s new ability to control your whole computer: keyboard, mouse, screen, any app. Felix Rieseberg pitches it as especially powerful with Dispatch so you can remotely control Claude from your phone, but the creator keeps hammering the obvious risk: if an AI with bad instructions gets loose on a real desktop with Chrome, email, and personal accounts, permission prompts are not going to save you.
“The Computer Will Use Itself” Meets a More Skeptical Reality
Dean W. Ball’s vision is that computers increasingly operate themselves while humans supervise like they’re playing a weird video game. The creator partially agrees, but pushes back hard on the idea that natural-language delegation replaces interfaces wholesale: shortcuts, macros, and purpose-built tools still compress intent better than asking an assistant every single time, even a very good one.
OpenAI’s Super-App Ambitions and Anthropic’s Shipping Spree
The video then pivots to OpenAI planning to merge ChatGPT, Codex, and Atlas into a desktop “super app,” which the creator finds partly sensible for coding and productivity, but less clearly justified around Atlas. That tees up a rapid-fire rundown of Anthropic’s 2026 launch velocity — Co-work, Opus 4.6, Sonnet 4.6, PowerPoint and Excel integrations, 1M context, Dispatch, projects, computer use, Auto Mode — capped with the running joke that there is basically a new Claude update every day.
Small Features, Big Workflow Upgrades
A bunch of lower-level improvements get called out as the kind of practical wins people actually feel: /BTW side conversations while Claude keeps working, Codex sub-agents, seven-day Claude Code loops, project imports, custom environments, remote Claude Code sessions launched from your phone, and persistent Co-work threads you can query later. The throughline is that these aren’t flashy benchmark announcements — they make agent workflows less annoying and more continuous.
Code Review Becomes a Real Product Category Fight
Anthropic’s new Claude Code Review gets one of the strongest endorsements in the transcript. Anthropic says large PRs average 7.5 issues found with under 1% false positives, Jared Sumner says it catches subtle bugs others miss, and the creator’s main objection is price, not quality: $15–$25 per review versus Codex review often costing around $1, which turns it into a classic quality-versus-cost tradeoff.
People Are Actually Building With This Stuff Now
The back half turns into a parade of examples showing agentic coding moving from hype to lived workflow. Gary Tan shares GStack, a slash-command skill stack for Claude Code; Bartosz Ciechanowski-like examples include turning audio into a full Jupyter notebook; Rob Miles says he tried replacing a bad app with Claude “and tragically it worked great”; and Trevor Lasn nails the vibe: this is how you accidentally build a startup.
The Craziest Part Might Be How Casual It’s Becoming
By the end, the creator is recapping people running long autonomous coding sessions, checking agents from their phones between gym sets, and increasingly not opening their editors because asking Claude is faster than reading code. There’s still a sharp warning not to confuse this with the reckless OpenClaw moment — Ethan Mollick says Co-work plus Dispatch covers 90% of what he wanted there while feeling much less likely to upload his whole drive to malware — but the overall message is unmistakable: stick to coding agents, because those are already becoming normal.