X / Twitter
Anthropic's Official Claude Account
Claude Code's desktop app has received a major redesign. Users can now run multiple Claude Code sessions side by side in one window and manage them through a new sidebar. The release also adds an integrated terminal, file editing, HTML and PDF previews, a faster diff viewer, and layouts customizable by dragging. CLI plugins are fully compatible with the new desktop app. Download or update the Claude desktop app to get it.
https://x.com/claudeai/status/2044131493966909862
https://x.com/claudeai/status/2044131495296586041
Anthropic Researcher Alex Albert
Albert confirmed the release of Claude Code's new desktop app, with extensive new features and performance improvements. In his experience, combining Cowork and Code means he almost never needs to open another app, or even a terminal, in his daily work. It is a practical endorsement of Claude Code's completeness as a full working environment.
https://x.com/alexalbert__/status/2044144091395699055
Claude Code Team Member Cat Wu
Claude Code introduced Routines, which can trigger Claude Code tasks on a schedule, through GitHub events, or through other APIs to automate repetitive daily work. This extends its use from manual invocation to continuous background automation.
https://x.com/_catwu/status/2044103103591657941
Vercel CEO Guillermo Rauch
Rauch officially open-sourced Open Agents, a reference platform for cloud coding agents. He explained why companies including Stripe (Minions), Ramp (Inspect), Spotify (Honk), and Block (Goose) are building their own AI software factories. Technically, general-purpose coding agents struggle in large monorepos and lack the company's institutional knowledge and custom workflows. Commercially, software companies' competitive advantage will shift from “the code they write” to “how they produce code.” Open Agents uses Vercel's Fluid for the agent brain, Workflow for long-running task durability, Sandbox for secure code execution, and AI Gateway for token routing across models. Its SDK can be used directly even without deploying on Vercel. He also recommended pairing DSQL, a Postgres-compatible elastic database service, with agent coding platforms.
https://x.com/rauchg/status/2043869656931529034
https://x.com/rauchg/status/2044067044325503404
Box CEO Aaron Levie
Levie described a new role he expects to become widespread in enterprises: deploying and managing agents. This person identifies the team's highest-impact workflows, existing or new; determines which tasks agents could do 100 times faster or at 100 times the volume; connects structured and unstructured data flows; designs points of human-machine interaction; and continually manages agent operations and KPI tracking. The role requires both business process knowledge and technical ability, including fluency with MCP, CLIs, and similar tools. Levie sees it as a distributed role, with every team needing one or more such people, and an ideal entry point for younger people to engage deeply with AI.
https://x.com/levie/status/2043883641366032638
Y Combinator CEO Garry Tan
Garry Tan released GBrain v0.9.3, adding search tuning, search evaluations, CJK query support for Chinese, Japanese, and Korean characters, better health checks, and several security fixes. He clarified that GBrain is his personal OpenClaw/Hermes agent configuration, with more than search optimization: it integrates multiple skill packages, including a voice agent built on OpenAI Realtime that lets Claw talk directly with users. It will next be upgraded to Gemini Live.
https://x.com/garrytan/status/2043948627291451625
https://x.com/garrytan/status/2043948971568312473
Cursor Designer Ryo Lu
Ryo Lu previewed upcoming Cursor support for split screens, customizable layouts above, below, left, and right, and multiple-agent views, allowing users to arrange the workspace their own way. This direction suggests Cursor is expanding into more complex workflows involving collaboration among multiple agents.
https://x.com/ryolu_/status/2043891602381500583
Replit CEO Amjad Masad
Masad offered an observation from a philosophy-of-technology perspective: most people are lax about measurable things but demand precision from inherently ambiguous ones. “Twisting data into the story you want to tell while demanding precision from things that cannot stand up in reality is merely performing rigor.” He also announced that Replit users in India can now build and pay through Razorpay, an important step in localizing Replit's payment ecosystem in India.
https://x.com/amasad/status/2043948524908490953
https://x.com/amasad/status/2044161522415612096
Google Labs VP Josh Woodward
Gemini added practice questions for India's NEET, the national medical entrance examination, as the country's biggest medical exam approaches. Woodward said the team is adding practice questions for more subjects and countries and publicly invited suggestions for what to add next. This reflects Gemini's continued investment in localized educational applications.
https://x.com/joshwoodward/status/2044149482548580689
OpenClaw Founder Peter Steinberger
Steinberger shared that OpenClaw had shipped a release entirely without his involvement, handled independently by the maintainers, which made him “unreasonably happy.” He said he was spending today and tomorrow in Vancouver, Canada, preparing for a TED Talks presentation, before returning to OpenClaw development.
https://x.com/steipete/status/2044047222481019300
Every CEO Dan Shipper
Dan Shipper released Sparkle v4, a Mac file-organizing agent. It inspects your file system and suggests a personalized organization structure based on how you work. It starts with a deep cleanup of old installers, duplicate files, forgotten screenshots, and system junk, then keeps things tidy in the background, with scheduled runs available. Sparkle is included in an Every subscription. Shipper called it “quietly one of the best apps” in Every's portfolio.
https://x.com/danshipper/status/2044079255726838273
Roblox Product Manager Peter Yang
Yang shared a trap he frequently has to remind himself about: with tools such as OpenClaw and Claude Code, it is easy to spend hours optimizing tool configurations instead of doing actual work with them. Configuration feels like progress and is fun, but beyond a certain point it no longer improves productivity. He also cited Figma CEO Dylan Field: “Don't treat AI output as the final result. You need to shape it like clay; you are the one who judges its quality.” Yang's conclusion: AI gets you to average quickly, but your taste takes you beyond average.
https://x.com/petergyang/status/2043848480423325757
https://x.com/petergyang/status/2044061039671820406
Builder Zara Zhang
Zara Zhang released a substantially simplified version of her Tab Out Chrome extension. It went from requiring a server and Node.js to a pure Chrome extension with no server dependencies. Saved tabs live in chrome.storage.local, with data kept 100% local. Her development principle: she spent 20% of the time adding features to this small app and 80% removing them. “Claude tends to overengineer things; you have to keep reining it in.”
https://x.com/zarazhangrui/status/2044114136364397043
Linear Head of Product Nan Yu
Nan Yu pushed back on frequent criticism of Google and Meta: these companies are enormous, and even their basic units of execution are large projects, so any mistake looks big. The right approach is to count wins rather than losses. It remains a perspective worth considering amid the AI wave, especially as Google's Gemini and Meta's AI investments continue to deliver results.
https://x.com/thenanyu/status/2044089885842186498
Podcasts
Training Data — From SEO to Agent-Led Growth: Profound's James Cadwallader
Key takeaway: Marketing's underlying logic has flipped. Your content is now written for agents to consume, rather than for people to read.
ProFound is a marketing platform for the AI era serving roughly 12% of the Fortune 500. It helps brands track and optimize their rankings and visibility in AI search engines such as ChatGPT, Claude, and Gemini. In this episode of Training Data, co-founder and CEO James Cadwallader systematically examines the logic behind Agent-Led Growth (ALG).
In traditional SEO, marketing content addressed people. Now its real consumers have quietly shifted to agents. When a user asks ChatGPT a question, it effectively sends an agent to crawl the internet, synthesize information, and generate an answer. That agent visits far more pages than an ordinary user. Cadwallader described using ChatGPT to find a showerhead: the system read 65 pages in one go. Agents' information intake and coverage of the long tail are therefore orders of magnitude greater than humans', rendering the traditional SEO logic of “reach the top five and you've won” ineffective here.
Citation sources differ substantially across AI platforms. Gemini relies heavily on YouTube because Google owns it. ChatGPT cites Reddit extensively for consumer products and favors LinkedIn for B2B. Claude's behavior changed noticeably between versions 4.5 and 4.6: its classifier became more sensitive and began supplementing its answers with live web searches more often. These differences mean brands can no longer use one content strategy everywhere; they must optimize for each platform's citation preferences.
Cadwallader rejects the idea that content in the AI era is necessarily “slop.” He cited research finding that 53% of readers preferred AI-written content in a blind test. In his view, the key is whether you can “tell an all-knowing superintelligence something it doesn't already know,” regardless of whether AI wrote the content. The concrete strategy is to offer original, firsthand insights, making the brand a more authoritative source than any existing public information.
He also identified a currently effective but unsustainable tactic: comparative listicles. AI models favor content that has already done the comparison work, as it offers the “path of least resistance.” For example, Sequoia could publish a list of “Silicon Valley's 10 Best VCs” and put itself first. This works in the short term, but Cadwallader predicts that stronger model reasoning will increasingly detect and downrank such manipulation.
He considers the “death of the internet” entirely possible within three years. If people stop clicking through to websites, advertising revenue collapses and the economic incentive to publish quality content disappears, leaving AI without high-quality firsthand sources. He expects AI labs eventually to vertically integrate social platforms, as with X + Grok, because social media is “the last bastion of authentic human signals.” Reddit's and Meta's closed-loop strategies already support this direction.
Cadwallader recommends relying on data rather than intuition: brands must use tools such as ProFound to track their actual visibility and citation sources across platforms, or they are merely guessing. Marketing teams must also use agents to do marketing, beyond automating their old tasks. Agents can integrate Gong sales recordings in real time, extract customer objections, and generate battle cards. Work that required a manual quarterly process 18 months ago can now run every day in real time.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8