X / Twitter
Replit CEO Amjad Masad
Amjad Masad wished people a happy Eid and shared a striking demo: connect Granola MCP to Replit, and Replit automatically starts building an MVP prototype as a product requirements meeting ends. He emphasized that Replit has been in the App Store for four years with the same core behavior: users enter or generate code, the server compiles it, and a WebView opens, without platform restrictions. This consistency lets Replit focus on improving AI capabilities rather than negotiating platform compliance.
https://x.com/amasad/status/2034877335917519068
https://x.com/amasad/status/2035077510720102465
https://x.com/amasad/status/2035091750533177648
Box CEO Aaron Levie
Aaron Levie argued that employees who effectively direct AI agents will have access to much larger compute budgets than ordinary peers in the same roles, with profound organizational consequences. He considers this one of the most important underlying structural changes, with the gap continuing to widen as models improve.
https://x.com/levie/status/2034833274506281061
Y Combinator CEO Garry Tan
Garry Tan shared a YC Easter egg he added to Claude Code’s `/office-hours` skill, reflecting his continued investment in AI coding tools. He also argued that Markdown combined with LLMs has effectively become a programming language, far better than traditional deterministic languages for high-value, complex tasks requiring contextual reasoning. He thanked Paul Graham, Jessica Livingston, and others for supporting his appointment as YC President and CEO.
https://x.com/garrytan/status/2034853668739252419
https://x.com/garrytan/status/2034881003140129145
https://x.com/garrytan/status/2034883862460080542
Vercel CEO Guillermo Rauch
Guillermo Rauch announced that Next.js 16.2 is explicitly designed for agent-native development. New AGENTS.md files and bundled documentation give AI agents precise knowledge of the Next.js version used by a project. The release also introduces `@vercel/next-browser`, dedicated browser tooling for agents that substantially improves their precision within the framework.
https://x.com/rauchg/status/2035076089861857500
Every CEO Dan Shipper
Dan Shipper joked about tax season by handing his tax documents directly to Codex. He also described the viral growth of Proof, an agent-native document editor: more than 4000 documents were created within two days of launch, but servers then began crashing frequently—and the whole product had itself been vibe-coded. He said stability fixes were underway and half-jokingly asked users not to use Proof temporarily to avoid overwhelming the servers again.
https://x.com/danshipper/status/2034996532944978315
https://x.com/danshipper/status/2035030133413704154
https://x.com/danshipper/status/2035042300871811547
Claude Code engineer Thariq
Thariq posted Eid greetings and shared his celebratory mood after a month of fasting and reflection. He also reposted Claude Code-related content showing agents in practical use, attracting broad community attention.
https://x.com/trq212/status/2035041983618830815
https://x.com/trq212/status/2035086678977912994
Claude Code operations team member Cat Wu
Cat Wu compiled four key principles for building agentic systems from frontline experience shared by Decagon product director Bihan Jiang and Datadog senior PM Kai Xin Tai: plan in short sprints; prioritize demos and evaluations over documentation; revisit features shelved for insufficient model capabilities after every model update; and always choose the simplest viable approach, because failures accumulate exponentially with complexity in agentic systems.
https://x.com/_catwu/status/2035104391137784111
https://x.com/_catwu/status/2035104389808222383
https://x.com/_catwu/status/2035104388541473275
Anthropic’s official Claude account
Claude announced Projects support in Cowork mode in the desktop app. Tasks and context can be centralized in one workspace, files and instructions stay on the local computer, and existing projects can be imported with one click. The update makes Claude Code use in team collaboration more structured.
https://x.com/claudeai/status/2035025494677365113
https://x.com/claudeai/status/2035025492617961704
Google Labs VP Josh Woodward
Josh Woodward announced a fix for a Gemini Android frustration: Gemini no longer cuts off recording prematurely when users pause during voice conversations. The iOS version will follow within weeks. The change makes voice interactions more natural and reduces frustration from being interrupted during pauses.
https://x.com/joshwoodward/status/2034797067344998862
Roblox product manager Peter Yang
Peter Yang sees OpenAI’s strategy becoming clearer: ChatGPT is the AI product with the largest installed base, and the next step is an all-purpose entry point for coding, knowledge work, and personal assistance. He believes “Research preview” is becoming the new “beta” label as iteration accelerates. He also summarized four lessons for agentic product development: short sprints, demos and evals over documents, revisiting old features as new models arrive, and prioritizing simplicity.
https://x.com/petergyang/status/2034784797168685322
https://x.com/petergyang/status/2034786020160635031
https://x.com/petergyang/status/2035130974682718585
Podcasts
Latent Space — Dreamer: the Agent OS for Everyone — David Singleton
This Latent Space episode features Dreamer co-founder and CEO David Singleton, introducing the AI agent platform for ordinary consumers in depth. Formerly Stripe’s CTO and an early contributor to Android at Google, David positions Dreamer as “Android for the agent era”: an open ecosystem where anyone can discover, build, and use AI agents.
At Dreamer’s core is a personal agent called Sidekick, which users can name and give a personality. It is both an everyday assistant and the platform’s kernel, analogous to an operating system kernel. Whenever one agent calls another, Sidekick must mediate to ensure data security and permission boundaries. This design gives Dreamer operating-system-level privacy and security protections.
The platform has three layers: Tools, Gallery, and Sidekick. Tools are open to third-party developers and include high-quality data interfaces for Google Search, Gmail, real-time Formula 1 data, MLB, and NFL, plus integrations such as Instacart and Granola. Tool publishers earn usage-based revenue shares, and the platform has established a $10,000 best-tool prize with a mid-April deadline. The Gallery already contains hundreds of community-built agents for personal productivity, email management, financial analysis, meeting notes, and more.
For the building experience, David demonstrated using Sidekick to create an AI Engineer conference app in 25 minutes, supporting speaker search, personalized schedules, and LLM-based recommendations for each time slot. Dreamer lets users share apps with people outside the platform, lowering distribution barriers. The underlying stack uses TypeScript, and code can be exported locally for editing with Cursor or Claude Code. The platform includes version control, though not Git, SQLite, multi-user permission isolation, and full agent execution logs.
On memory, David said the team had tried vector databases, RAG, knowledge graphs, and other approaches. Its current system resembles those of mainstream products but still has dedicated ongoing optimization, aiming for Dreamer to understand users better over time.
David also introduced Builder in Residence, recruiting creative outside builders with immediate access, opportunities to work directly with the team, and paid agent-building opportunities. The company has around 17 people and maintains high talent density. Hiring emphasizes candidates’ ability to collaborate with coding agents, with a practical interview exercise building a complete product feature alongside Codex or Claude Code.
https://youtube.com/watch?v=TvmxWWfiYWI
No Priors — Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
This No Priors conversation with Andrej Karpathy covers code agents, automated AI research, changing engineering roles, robotics, education, and other topics.
Karpathy described entering “AI psychosis” in December 2024: his coding shifted from 80% written himself and 20% delegated to agents to almost never typing a line manually. He considers this an epochal productivity leap that ordinary engineers and outsiders have barely noticed. His current workflow runs multiple agent sessions, reviewing one agent’s output while another works in an efficient round-robin loop.
On AutoResearch, Karpathy described a concept he is exploring: agents automatically search for commits that optimize training code. Verification is cheap, requiring one training run to check validation loss, while the search space is enormous. He also envisions a decentralized research framework resembling a blockchain: untrusted workers on the internet submit candidate commits, anyone cheaply verifies their effectiveness, the proof of work is large-scale experimentation, and the reward is leaderboard position rather than money.
On engineering roles, Karpathy believes “coder” is no longer accurate; “expressing intent to agents” is closer. As agents do more, engineers’ real value concentrates in the few key decisions agents cannot make, such as taste in product direction and identifying core abstractions. He cited microGPT: agents can understand and implement it, but cannot independently conceive that minimal abstraction.
In education, Karpathy proposed a new paradigm: future courses and learning materials should be designed for agents rather than people, replacing HTML pages with Markdown documents that agents digest and explain to humans. He believes excellent educators will become people who “inject key cognitive bits into agents,” while agents handle other knowledge transmission.