Quick Hits — August 13, 2026
A midday batch: Anthropic maps what happens when agents talk to agents, Claude helps close a 30-year-old math gap, and a Chinese AI workbench turns generated webpages into living documents.
Anthropic's red team published a report on emerging multiagent systems — the lab's first serious attempt to map what happens when AI agents start interacting with each other at scale. It argues agent-agent traffic could plausibly outpace human-human and human-agent interaction before institutions adapt, and shows how benign quirks like confabulation and reward hacking compound into systemic failures in shared codebases, markets and social systems — in one experiment, 45 agents with individual virtual machines and a shared coordination forum were let loose hunting vulnerabilities across 15 open-source projects. The take: institutions built around human-speed oversight are about to be stress-tested by peers that never sleep.
Anthropic mathematician Levent Alpöge says he and a team of three humans plus Claude have constructed the order-668 Hadamard matrix, closing a 30-year gap. Alpöge posted the result as a 23,828-character string of plus and minus signs on X; decoded, it contained all 12 remaining Hadamard orders below 2,000, each re-verified with exact integer arithmetic, and Epoch AI has provisionally marked the FrontierMath open problem solved — the fourth of 50 to fall to AI. The take: it reads as clever search plus human pattern-spotting, and the pending technical report will decide whether the method generalizes.
WorkBuddy's 5.3.11 update turns AI-generated webpages into persistent, editable "light apps" — pages backed by live CSV storage that people and agents can co-edit, then publish to a shareable link in one click. The upgrade extends the Chinese work assistant's human-AI co-writing from Office documents to HTML and Markdown, so editing a generated page now works like editing a Word document: select anything, ask for a change in plain language, accept or reject AI-suggested revisions, and watch multi-user edits sync in real time. In a hands-on test, Zhidx built a working operations dashboard from a single prompt with no code on the free Hunyuan 3 tier — the take being that AI-generated pages are becoming tools you keep, not posters you screenshot.
