Issue 1: memory everywhere, agents in the browser
The pilot issue: Claude’s memory now spans chat, Cowork and the cloud, Cowork agents can open and act on web pages, and the API platform went generally available.
- Where the money actually runsRamp spend data via the FT: Fable 5 at 8% of Anthropic spend, mid-tier Opus 4.8 at 28%. Production runs on cheap.
- Anthropic funds wellbeing-impact evaluationsPolicy side: grants for measuring how AI actually affects people.
- A million agent-written lines, running in productionPaul Dix, quoted by Simon Willison: AI wrote a million lines and refined them over months into software now on millions of developer machines. The frank version of the number everyone keeps guessing at.
- Stop Making TUIsAgents made real GUIs cheap; the terminal is no longer the default.
- Screen memory without screenshotsPlain-text screen context an agent can read, cheaper than screenshots. A 43-point HN thread.
- Claude's memory now works everywhereMemory spans chat, Cowork and the cloud, with every stored topic visible and editable in Settings. The assistant stops being surface-local: the week's real shift.
- Cowork agents get a built-in browserAgents open and act on web pages directly, and Claude in Chrome went generally available the same week. The read-the-web layer stops being the bottleneck.
- The API platform went stable in one waveComputer use out of beta, a browser-use tool that drives a hosted browser from the accessibility tree, and the production-agents guide walking computer use, Skills and Files through one build.