## 🔍 Core Insights This week the AI world shows a clear trend of "de-bubbling and focusing on real deployment": **Anthropic** recommends streamlining subagent architecture, and the industry is beginning to reconsider the necessity of complex multi-agent systems—a single agent combined with **skills** can instead deliver roughly a **2x** execution speedup [0]. Meanwhile, **Google** CEO **Sundar Pichai** has already used internal agents to automatically summarize training progress, compressing an hour of daily follow-up [12], while **Block** released **Dessa**, a full-duplex speech model that can run locally on a single GPU [18], showing that "small and refined" and "real deployment" are becoming the new main themes. ## 🚀 Key Updates - **Dex Horthy on Anthropic's recommendation to streamline subagents** [0]: A single agent combined with skills often beats multi-agent setups, with customer execution speed improving by roughly 2x. - **Krea opens Krea 3 image editing beta** [3]: End-to-end 2k editing completed in 8 seconds, supporting bounding box control and precise multi-turn editing. - **Sundar Pichai on using AI agents to manage Google** [12]: Agents automatically summarize pretraining and post-training progress, compressing daily follow-up time. - **Block releases local full-duplex speech model Dessa** [18]: Runs locally on a single NVIDIA GPU, can speak and listen simultaneously, with emotional expressions such as laughter and pauses. - **Perplexity launches open-source skill /impeccable** [19]: Checks layout, fonts, contrast, and accessibility, with 5,000 credits gifted per user. - **Kimi K3.1 Pro (Max) ranks 11th on Agent Arena** [23]: Achieves a +7.6% net improvement based on 4,000+ real conversations, with open weights coming soon. - **Pennsylvania signs AI deceptive advertising law** [11]: Ads generated or significantly modified by AI must carry a label in the same format, applicable to any ad run in Pennsylvania. - **Three fired OpenAI safety researchers speak out** [21]: They deny sharing confidential information and say they were never informed in writing of the alleged leaked content. ## 🔗 Sources [0] Dex Horthy on Anthropic's recommendation to streamline subagents: A single agent plus skills is often better — https://aihot.news/items/efeqm1hmbbfxtlvpyp90syozs [3] Krea opens Krea 3 image editing beta, supporting fast 2k generation and precise multi-turn editing — https://aihot.news/items/xougitbm9f5vrlfghlw4gh7t6 [11] Pennsylvania signs bill requiring AI-generated ads to carry labels — https://aihot.news/items/k50ff2029wwvy2pinbmoceo3d [12] Sundar Pichai on how to use AI agents to manage Google — https://aihot.news/items/f9lp2z2y29vunz0j8lrt69ja6 [18] Block releases Dessa, a locally run full-duplex speech model — https://aihot.news/items/m12zx4vkj1e4avkq0na431zbf [19] Perplexity launches open-source skill /impeccable on the Computer skills marketplace, gifting 5,000 credits per user — https://aihot.news/items/jryc594pqpq0xbtqbbd6yn7xb [21] Three fired OpenAI safety researchers say they were never told what was leaked — https://aihot.news/items/b2o0