New v0 API enables programmatic app building

#1 𝕏

A new v0 API was introduced, providing programmatic access to v0's app-building capabilities, including starting a chat from a prompt, repository, or ZIP file, rendering a development server preview, sending follow-up messages, and deploying to Vercel.

#2 📝 OpenAI News

Continuous voice interaction with GPT Live - An Engineering post introducing continuous voice interaction capabilities for GPT Live, describing technical work and user experience improvements to enable ongoing voice conversations. The article explains engineering approaches and expected benefits for real-time voice use.

Also covered by: @OpenAI

#3 𝕏

Cursor announced new plugins that let agents read, write, and act across Google Workspace, with direct access to Gmail, Google Drive, Calendar, Docs, and Sheets.

#4 𝕏

OpenAI announced that an internal version of its next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates.

#5 𝕏

LlamaIndex 🦙 announced that LiteParse can extract structured PDF data—including form field values, checkbox states, annotations, embedded images, vector graphics, tagged document structure, and word-level bounding boxes—in milliseconds per page. New complexity signals cover scanned pages, multi-column text, ruled or borderless tables, and dense figures, helping route pages that need a model to tools such as LlamaParse.

#6 𝕏

Guillermo Rauch recapped nextjs.org’s Next.js 16.3, highlighting faster development, builds and type-checking, an incremental build cache, improved memory use, and more cost-efficient serving. It also brings opt-in instant navigations planned as the default, Suspense fallback choices and versioned docs for agents; Rauch said an agent migrated his projects without issues and thanked the release’s 90 contributors.

#7 𝕏

NVIDIA AI shared an AI Model Co-Design post explaining why long-context serving speed is largely determined before training. It highlights four architecture choices—group size, head dimension, KV-cache size, and parallelism strategy—that shape system throughput and per-user responsiveness.

#8 𝕏

Cursor announced that its Cloud agents are now 20–30% more token efficient and 80% more efficient on runs involving computer use. Improved handling of MCPs, skills, and computer use supports delegating more ambitious tasks and receiving demos while staying within budget.

#9 ▶️

How this OpenAI engineer uses Codex + ChatGPT Work to automate everything

How I AI Podcast

Also covered by: @claire vo

#10 𝕏

Guillermo Rauch demonstrated basic browser automation in Vercel Sandbox using the remote-agent-browser API to open a URL, click, and take a screenshot.

#11 𝕏

claire vo recommends @evedev_ as a default framework for internal agents, citing its instructions, skills, built-in channels, and connectors. She also announced a forthcoming How I AI episode about building a PR review and approval agent with Eve, while a quoted post describes Vercel’s internal AI agent @v as powered by @evedev_.

#12 in

Greg Isenberg shared a recommendation that startups use an agent to update a daily “what_the_market_is_telling_us. md” file every morning, drawing on customer and market signals across payments, product analytics, support, calls, CRM, issue tracking, and outside research. The file should surface changing patterns, supporting receipts, and potential product or GTM decisions—not merely summarize events.

#13 𝕏

Thariq said that when a Claude Connector is connected—such as Gmail, calendar, or Slack—Claude Code can also use it, including in Artifacts.

#14 ▶️

Why Graph Engineering will 10x your Claude/Codex

Greg Isenberg

#15 in

Peter Yang recapped six takeaways from Nous Research co-founder Karan Malhotra on using Hermes as a personal agent, including building personalization through memory and skills, switching response styles, and using one agent to work and a fresh agent to evaluate. Malhotra also highlighted skill cleanup, local and model-flexible use, preserving open source, and using Hermes for creative projects.

Also covered by: @Peter Yang

#16 in

Marc Baselga recapped a Training Data podcast episode featuring Katelyn Lesse and Angela Jiang, who lead engineering and product for Anthropic’s Claude Developer Platform. He describes Anthropic’s strategy as open to shared standards and third-party agent sandboxes, but opinionated about model-specific harnesses—a choice he argues could make switching from Claude require significant agent retuning.

#17 in

Udi Menkes recapped an AI architecture that connects general knowledge with documented real-world experience to produce domain-specific judgment. At Intuit, frontier models generate hypotheses for actions a business can take. A model trained on outcomes selects actions that may help, drawing on millions of business trajectories from QuickBooks, TurboTax, Credit Karma and Mailchimp that link a state, an action and an outcome; a language model then turns the selection into a clear recommendation.

#18 𝕏

Rowan Cheung shared “The project memory system,” which uses one authority document, versioned files, decision logs, and a handoff prompt to preserve context across new chats in long AI projects. Cheung credited Christian Denton.

#19 𝕏

Santiago demonstrated how to set up a web research agent in 5 minutes that runs 24/7 and returns structured search results, with instructions for configuring source limits, goals, output format, and compute. The post, created in partnership with @nimble_search, includes a video demo and a link to try the agent.

#20 𝕏

Qwen3.8-Max is available to try on Venice, where users can ask questions anonymously.