Alexandr Wang, head of Meta's Super AI Lab (MSL), announced the launch of Muse Image, claimed to be the first 「agent-style」 image generation model. Its key feature is performing reasoning and planning before generating images, rather than producing outputs directly. Meta also previewed its companion video generation model, Muse Video, highlighting competitive performance in prompt adherence, visual fidelity, and temporal consistency. This marks a paradigm shift in image generation—from 「one-shot direct output」 to 「think first, generate later」—diverging from the immediate generation path of traditional diffusion models. Currently, only product announcement details are available; specific parameters, availability timeline, and pricing have not been disclosed.

Google has introduced four new features for the managed agents in the Gemini API: background task execution, support for remote MCP servers, function calling, and credential refresh—all now accessible via the free tier. Logan Kilpatrick, Google’s developer relations lead, stated the goal is to significantly reduce the cost and friction of deploying agents into production. The service is already used by thousands of customers, with future roadmap items including a new UI and custom agents. This move positions Google to compete with LangChain, Anthropic, and others in the agent infrastructure space.

Apple and Broadcom have extended their partnership through 2031, jointly developing custom ASIC chips for multiple future product generations to support Apple’s planned deployment of advanced AI servers in 2027. This reflects Apple’s strategic push to build proprietary AI computing infrastructure, securing stable chip supply through deep collaboration with Broadcom in preparation for large-scale AI service rollouts. This long-term agreement underscores the industry trend of major tech firms designing their own chips to reduce reliance on NVIDIA.

NVIDIA has delayed the release of its Kyber rack-scale architecture by more than 12 months, pushing it to 2028. Designed to integrate 144 Rubin Ultra chips into supercomputers, the delay stems primarily from manufacturing challenges with core circuit boards. This highlights how NVIDIA’s aggressive iteration pace is being constrained by real-world fabrication bottlenecks. Separately, Apple’s upcoming iPhone Ultra, expected in September, is projected to price at $2,400—double that of the iPhone 17 Pro Max—potentially triggering high demand and supply shortages.

LangChain co-founder Harrison Chase announced the open-sourcing of the deepagents project, a model-agnostic agent runtime framework designed for long-running workflows such as research and coding, capable of handling planning, context management, and multi-agent orchestration. Concurrently, LangChain Academy has launched a free accompanying course covering framework concepts, core capabilities, building, and deployment, with native integration of LangSmith for tracing and deployment. LangChain emphasizes that real-world AI applications almost always require a software harness around LLMs to connect tools and external environments.

Anthropic announced that Claude Cowork will expand to mobile and web platforms, allowing users to hand off tasks across devices and receive completed work. The test version will roll out over the coming weeks starting with Max subscribers. New capabilities include scheduled tasks that run even when computers are off (e.g., automatically preparing briefings or drafting follow-up emails), and unified interfaces integrating Chat and Cowork across web and desktop with shared projects and artifacts. Additionally, doubled Cowork usage limits have been extended through August 5. Cowork is positioned as an active business assistant, not a passive chatbot.

Anthropic’s ClaudeDevs shared benchmark results on cost-optimized configurations of its managed agents. On the BrowseComp benchmark, using Fable 5 as orchestrator and Sonnet 5 as worker achieved 96% of Fable 5’s full performance at 46% of the cost. On SWE-bench Pro, a setup where Fable 5 acts as advisor and Sonnet 5 as executor reached 92% performance at 63% cost. This approach leverages Fable 5 for planning and Sonnet 5 for execution, with caching of sub-agents to avoid full token costs from repeated calls. Anthropic has also extended the free trial of Fable 5 until July 12.

Google DeepMind has launched a new skill called 「Predicting the Past」 on its Antigravity platform, integrating Gemini with expert AI models Aeneas and Ithaca. This enables historians to study ancient Greek and Latin texts using plain English, eliminating the need for specialized technical expertise. In collaboration with historian Thea Sommerschield, DeepMind completed three case studies validating the skill’s research value in text restoration, dating, and authorship attribution, demonstrating concrete pathways for AI application in humanities and historical research.

A practical implementation shows that using small language models to prune context in Retrieval-Augmented Generation (RAG) systems successfully removed 68% of irrelevant content while preserving 96% recall, significantly improving question-answering efficiency. The core insight is that not all tasks should rely on large LLMs—deterministic tasks should be handled by application-layer logic to save costs. The article also notes that the emergence of efficient small models like GLM 5.2 may compress profit margins in AI services and predicts that AI labs will spend over $100 billion annually on data by 2030, underscoring data as a strategic resource.

Schneider Electric has deployed more than 60 AI agents in global production environments, serving 160,000 employees, and uses self-hosted LangSmith for tracking and observability management. This case represents large-scale industrial enterprises bringing AI agents into real-world production. In parallel, DoorDash has rolled out Claude Code company-wide, observing that faster code generation shifts pressure onto CI/CD pipelines, code reviews, and security reviews—highlighting that AI-driven efficiency gains are a 「whole-company issue」 rather than just an engineering challenge.