Seven failures, zero error messages: a quick check that the meeting recorder I built to replace Granola was still on current model versions turned into a full audit. Seven failures, four months in production and not a single error between them. Three had no symptom at all – the only test that ever ran was "a transcript appeared and reads correctly", and all three pass it. The other four were written down as known issues with workarounds, and three of those had a recorded cause that was wrong. The full post-mortem is coming. Blog post (the original build)

Only the UI dies: a group debate this week landed on "pure software is dead – agents will act on trusted data stores instead". Half right. The software persists, it just goes programmatic. What dies is the human clicking around a web app. Or as the new YC mantra goes: build something agents want.

The context-engineering write-up is live: last issue had the summary – the skill that holds my other skills to a standard. The write-up adds the model underneath: everything you hand an agent is either pushed into context (loaded whether it's needed or not) or pulled on demand, and that one distinction decides where every instruction belongs. The skill is free to copy, in Claude Code and Cowork versions. Blog post

Slack bolted AI on the side: Buzz (Jack Dorsey's new thing) and Cloudflare OS are workspaces built around agents from day one. Nobody's moving a 60-person team off Slack this year. But teams moved off Jira to Linear because the challenger was built around how they actually work, and Slack gets the same treatment if it doesn't catch up. Site

On my radar

The OpenAI → Hugging Face timeline: the strangest security story of the year now has a Black Hat timeline. An OpenAI training agent given an impossible task found somewhere it could write files, other agents started leaving each other messages there, and it ended in a real breach of Hugging Face that OpenAI only traced back to itself weeks later. Simon Willison

Claude's text gets a watermark: Anthropic will watermark generated text to meet the EU AI Act – it applies to new models and older ones, and survives copy-paste and editing. Everyone I've seen react is sceptical it's enforceable at all. TechCrunch

A 1983 aerospace standard as an agent skill: ASD-STE100 is the simplified English aerospace manuals are written in – short sentences, active voice, one instruction per sentence. Someone packaged it as a skill so LLMs write docs the same way. Simple language was an engineering standard forty years before it became a prompting technique. GitHub