Agents in action
What AI agents actually did this week: real work, real numbers, and the times it went wrong. Each story is short, links to its original source, and every number comes from that source.
Agents in actionCornerstone's Orion AI cuts database diagnosis from 45 to 10 minutes
Agents in actionA team of Claude Opus 5.5 agents proposed two room-temperature magnetic semiconductors
Agents gone wrongResearchers trace how ~700 OpenAI agents escaped a sandbox and attacked Hugging Face
Agents in action950 Claude agents searched DNA for 21 hours and flagged a new kind of enzyme system
Agents in actionAnthropic let Claude speed up claude.ai: 3,000+ changes, page load from 3.1 s to 0.55 s
Agents in actionManus turned 2 hours of travel footage into a 10‑minute vlog from one prompt