Alibaba's DOM-only web agent, and design loops beat prompts

Alibaba's Page Agent drives web UIs through the DOM, no screenshots — a JavaScript in-page agent that reads the live DOM as text and clicks and types from natural-language commands. No vision model, no screenshot pipeline, no backend rewrite. If it holds up, that's a meaningful latency and cost cut versus Claude's Computer Use and the rest of the screenshot-first crop — worth a look for anyone doing RPA-shaped work. — Asif Razzaq

Ornith-1.0: self-improving open-source models for agentic coding — an open-source framework for agentic coding that claims a self-improvement loop on top of the base model. Whether the loop actually cuts error rates versus a static baseline is the only question that matters here; the repo is the place to find out. — danboarder

Design Loops, Not Prompts — makes the case that self-checking loops collapse into circular reasoning — the model happily validates its own hallucinations. Marin's fix: structured feedback loops with external validation. As he puts it, "don't let the model check itself." Sensible advice for anyone still one-shotting production prompts. — Javier Marin

DEEP DIVE

Using Local Coding Agents — Raschka walks through running open-weight models in local coding harnesses as a replacement for Claude Code and Codex subscriptions. The tradeoff is honest: no API fees, no egress, but you're bringing your own GPU and accepting a real accuracy gap on hard tasks. — Sebastian Raschka, PhD

shot-scraper 1.10 lets your agent record video demos of its work — adds Playwright-backed video recording driven by YAML storyboards. Willison's pitch: coding agents should produce demos of what they built, and now they can — without a human stitching together screen recordings. Handy for eval pipelines and PR artifacts. — simonwillison.net

For most of the world, open-source AI is the only way forward — argues that outside a handful of wealthy markets, proprietary API economics simply don't pencil out — leaving open weights as the default deployment path. A useful reminder for founders whose pricing decks assume everyone can afford frontier tokens. — CrankyBear

AI boom risks global financial crash, warn central bankers — central bankers add AI capex to the list of things that might blow up the financial system. Not immediately actionable, but if you're raising or planning big infra spend, this is the drumbeat regulators tighten to. — b-man

Europe Is Fed Up and Wants Its Own AI — Relevant for founders considering EU market entry or compliance: new funding initiatives and AI Act enforcement will reshape model deployment strategies in Europe over 2025. — wired.com