This week, AI agent capabilities made significant strides as Claude launched built-in memory for its Managed Agents and expanded its connector directory with popular apps. Google DeepMind focused on accelerating AI adoption through major consultancy partnerships and introduced new research for resilient LLM training. Additionally, Google Labs rolled out an AI-powered marketing tool for small businesses, while AI coding assistant Cursor secured key enterprise and infrastructure collaborations.
The Flow Sessions program recently completed its third chapter, uniting diverse creatives from journalism, advertising, and fashion to experiment with Flow beyond traditional filmmaking. Artists learned to embrace experimentation, like Julie Wieland using Flow as an "endless playground," and to infuse personal stories into their work, as demonstrated by Calvin Herbst's visual elegy. The program also showcased Flow's versatility across industries, with fashion designer Charline Prat creating exaggerated textures and journalist Chloe Desaulles exploring realism, ultimately highlighting Flow's ability to unlock new creative expressions and storytelling possibilities.
Read original →Claude has launched built-in memory for its Managed Agents, now available in public beta, allowing agents to learn and improve across sessions. This intelligence-optimized memory stores information as files on a filesystem, giving developers full programmatic control, exportability, and API management. Designed for enterprise, it features scoped permissions, audit logs, and allows multiple agents to share or access memory stores concurrently. This enhancement enables continuous learning, leading to fewer errors, faster task completion, and eliminating the need for custom memory infrastructure.
Read original →Claude is expanding its connector directory to integrate popular apps for everyday life, such as AllTrails, Instacart, Audible, TripAdvisor, and TurboTax, building on its existing work tool integrations. Users can now perform a wider range of tasks directly within Claude conversations, like finding a hike or ordering groceries, by connecting these services. Claude dynamically suggests relevant apps based on the conversation's context, streamlining interactions. The platform emphasizes user control over actions and data privacy, ensuring no data is used for model training and that users approve any purchases or bookings.
Read original →Decoupled DiLoCo is a novel distributed architecture designed to improve the resiliency and flexibility of training large language models across geographically distant data centers. It tackles the challenges of tightly coupled, synchronous systems by dividing training into asynchronous "islands" of compute, which isolates disruptions and allows continuous learning despite hardware failures. This self-healing infrastructure, building on advances in asynchronous data flow and low-communication, reduces bandwidth requirements and maintains training availability and performance even when learner units fail. Demonstrated with Gemma 4 models, Decoupled DiLoCo offers a more robust and scalable approach to developing frontier AI.
Read original →National Australia Bank (NAB) has standardized its AI coding assistant tools on Cursor for 6,000 developers, with plans to scale to over 10,000 employees. This decision followed an evaluation against Amazon Q and GitHub Copilot, with NAB choosing Cursor for its superior model flexibility, deep understanding of complex codebases, and extensibility. The adoption has led to significant productivity gains, including a threefold acceleration in legacy migrations like monolith refactors and mainframe transitions. Additionally, teams are now developing greenfield applications, such as a payment app, in just three weeks instead of the original four-month scope, enabling NAB to ship projects faster and with higher quality.
Read original →Anthropic's Economic Research team is launching the Anthropic Economic Index Survey, a new monthly qualitative survey conducted via Anthropic Interviewer. Starting April 22, 2026, the survey aims to capture how Claude users experience and anticipate AI's impact on their work, productivity, and the broader economy. This initiative seeks to provide crucial first-hand accounts and identify economic shifts as they emerge, complementing existing quantitative data and potentially surfacing changes before they appear in aggregate labor market statistics. A small, randomly selected group of Claude users will be invited each month, with insights to be published in future Anthropic Economic Index reports.
Read original →Connecting AI agents to external systems effectively is crucial for their utility, with teams typically using direct API calls, CLIs, or the Model Context Protocol (MCP). While direct API calls struggle with scalability and CLIs have limited reach, MCP provides a standardized common layer for integration, handling authentication, discovery, and rich semantics. This protocol is increasingly vital for production agents, which are often cloud-hosted and require robust, portable connections to remote systems. MCP is seeing significant adoption, with SDKs surpassing 300 million monthly downloads, underpinning major agentic platforms like Claude. To maximize effectiveness, building remote MCP servers is recommended for broad distribution across various client environments.
Read original →Google DeepMind is partnering with leading global consultancies—Accenture, Bain & Company, BCG, Deloitte, and McKinsey—to accelerate AI-driven transformation for businesses worldwide. This collaboration aims to bridge the significant AI adoption gap by combining Google DeepMind's advanced research and frontier AI models, such as Gemini, with the consultancies' strategic expertise. The initiative will focus on developing scaled, industry-specific AI solutions, providing early access to cutting-edge models for partner feedback, and connecting AI leadership with customer CEOs. Ultimately, this partnership seeks to help organizations responsibly harness AI's potential for economic growth and problem-solving across various sectors.
Read original →A survey of 81,000 Claude users explored the economic impact of AI, revealing that individuals in roles with higher AI exposure, particularly early-career professionals, expressed greater concerns about job displacement. Paradoxically, many respondents also reported increased productivity and empowerment, with the highest- and lowest-paid occupations experiencing the largest gains, often through expanding task scope. The study established a direct correlation between a job's observed AI exposure and its occupants' economic concerns about potential job loss.
Read original →Cursor has announced a partnership with SpaceX to accelerate its model training efforts. This collaboration aims to overcome Cursor's previous compute bottlenecks that have limited the development of its agentic coding models, such as Composer 2. By leveraging xAI's Colossus infrastructure, Cursor plans to dramatically scale the intelligence and capabilities of its AI models.
Read original →Google is rolling out its AI-powered marketing tool, Pomelli, to small businesses across the E.U., Iceland, Liechtenstein, Norway, Switzerland, and the U.K. Developed by Google Labs in partnership with DeepMind, Pomelli is designed to help businesses easily generate on-brand content like product images and social media campaigns. The tool works by analyzing a business's website to understand its brand identity, suggesting custom campaign ideas, and then creating high-quality, editable assets for social media, web, and ads. This expansion aims to empower small businesses with advanced AI technology for creating professional, consistent marketing materials.
Read original →Cursor has significantly reduced disruptive app crashes, primarily caused by out-of-memory (OOM) issues, which escalated with increased users and complex features like subagents. To address this, the company implemented robust systems for crash observability and memory pressure monitoring, focusing on severe renderer crashes. Their approach involves dual debugging strategies: a top-down investigation of memory-intensive features and a bottom-up tracing of individual crash events using real-time stack captures. These efforts have resulted in an 80% reduction in the OOM-per-session rate and a 73% drop in the OOM-per-request rate since late February/early March.
Read original →Stitch has open-sourced its DESIGN.md format, a specification that enables users to export and import design rules across projects. Initially, this allowed Stitch to understand design systems and generate brand-consistent user interfaces without constant re-invention. Now open-source, DESIGN.md can be adopted across various tools and platforms, empowering AI agents to accurately interpret design intent and validate choices against WCAG accessibility rules, thus fostering a shared visual language.
Read original →The "Built with Opus 4.6 Claude Code hackathon" announced its winners, showcasing diverse participants like a personal injury lawyer, cardiologist, and musician, many of whom were not professional developers. First place went to Mike Brown's CrossBeam, an AI-powered tool designed to automate and streamline California's complex housing permit application process. Another highlighted project, Elisa, by Jon McBee, creates a block-based visual IDE for Claude Code, making it accessible for non-developers. These innovative projects aim to inspire future builders for the upcoming Opus 4.7 Claude Code hackathon.
Read original →
