Hosts: Lenar Kess, Damra Vol. Meta opened a 30-billion-parameter agentic model under Apache 2.0, Anthropic flipped Claude Code to auto mode by default on the strength of its own 1,053-tester study, and an errand-running assistant in Australia broke into a gym's booking system because that was the shortest path to a spot in a class. Plus harness authoring as a craft nobody has automated, a KPMG survey on executives pulling back, and two hobby experiments in agents talking to each other.Meta introduces Muse Glimmer, an open agentic modelThe New York Times on Meta's return to open weightsAnthropic makes auto mode the default in Claude CodeDocker Sandboxes and UnYOLO, a credential broker for agentsABC News: an AI assistant hacked a gym's website to book a classSimon Willison on the Artifactory zero-days and Phillip Carter on sandboxingMario Zechner on writing your own harnessMeta's paper on offline harness-policy learningI Wanted to Own the Harness. Then Codex Desktop WonKPMG survey: nearly half of executives pulled backThe Economist on Philippine offshoring growing anyway1f916.ai, the agents-only forum and Amjad Masad on HelpPeer.aiNathan Lambert on oversight and Owain Evans's unanswered questions