AI Today · April 9, 2026Key highlights:- Meta launches Muse Spark, its first model from Superintelligence Labs with multi-agent Contemplating mode achieving 58% on Humanity's Last Exam- Meta consumed ~60 trillion Claude tokens in 30 days, raising questions about token economics and potential reasoning trace distillation for Muse Spark- Anthropic launches Managed Agents architecture decoupling brain, hands, and session for long-running agent work with 60-90% latency improvements- Google Colab introduces Custom Instructions and Learn Mode turning Gemini into a personal coding tutor- Agno founder argues agentic software needs systems engineering across five critical layers- PyTorch Monarch framework adds Kubernetes support, AWS EFA RDMA at 16 Gbps, and AMD ROCm GPU support- Claw-Eval benchmark provides 300 human-verified tasks with Pass^3 methodology for reliable AI agent evaluation- Cursor BugBot reaches ~80% bug resolution rate through self-improving learned rules, outperforming competitors by 15 points- Federal appeals court denies Anthropic's bid to block Pentagon blacklisting while separate court bars Claude usage ban- AI agent Poke raises $25M at $300M valuation for text-message-based automation accessible via iMessage, SMS, and Telegram