The cloud box write-up is out: A few people messaged after last week asking how to set it up, so I've written it up with the six build scripts for anyone who wants to copy it. The design call that mattered more than any of the parts is that the box is primary and nothing syncs. The tempting setup is laptop-first with a sync tool keeping the server in step, and that's exactly where sync tools break: agents editing uncommitted files on both sides at once. So the laptop keeps a checkout under one rule: git-backed work only, commit and push before switching back. If it isn't in git, it doesn't exist on the laptop. Blog post

Start a website rebuild from an Astro theme: Someone in a founders group was replacing their Squarespace site by having an agent drive the browser, and burning through their token allowance doing it. Computer use is context-heavy because every move captures the screen. The quicker path is to pick an Astro theme (a free framework for content sites – images, SEO and speed handled for you) and have the agent tweak it from there. That's what I did for my personal site and fidero.com, and paying for a theme you like is a no-brainer. Blog post (fidero.com rebuild)

The phone app is a window, not the engine: Claude's Remote Control puts a live session in the iOS app and is the one native feature that fits the box. But Anthropic's own docs tell you to run it inside tmux (a terminal session that lives on the server), because on its own it doesn't survive a dropped connection. It's a phone view on top of tmux. The hosted options (Claude Code on the web, routines) run on Anthropic's machines rather than yours, so they can't see your connectors or read across your repos. Blog post

The research got the architecture right and the specifics wrong: Six research agents produced the runbook the box was built from. Every architectural decision survived the build but almost none of the detail did: the host it chose had no capacity, the phone app it rejected was fine and eleven secrets wrote themselves blank without an error. That's the next post.

On my radar

Thoughtworks' CTO would rather not review all this code: Rachel Laycock's numbers: lines per human-landed diff at Meta up 106% in a year, DX's median pull request up 64%. Her answer is to move the judgement earlier (pairing, design sessions, fitness functions) and review by exception, and she calls an AI reviewer "automating the ceremony". I half agree. My agents' code passes four reviews before I see it, and what I read is the spec and the PR body, not the diff. Post

One model, 62.7% or 99.9%: ARC Prize ran OpenAI's new GPT-6 Astra on ARC-AGI-3 twice. With a neutral test setup it scored 62.7%. With OpenAI's own context management switched on (it keeps its reasoning state between turns and compacts long conversations) it scored 99.9%, faster and on 49% fewer tokens. The model was the same and the context handling wasn't. ARC Prize

OpenAI agents found a public wiki and started talking to each other: Researchers found about 18,000 posts by OpenAI agents on a German wiki, written during a web research task where writing to the internet was supposed to be blocked. They shared answers, and one posted a trick for getting round the sandbox's network limits. Another agent ran it 14 minutes later. Editing stopped the day after OpenAI staff IPs visited. Nightingale

Meta's transcription model knows who said what: Muse Voice Transcribe does live transcription with speaker separation for 20+ people, and ranked first on Artificial Analysis for streaming speech-to-text at launch. Speaker labelling was the exact thing that made me replace Granola. Meta