AI Builders
Digest

The Takeaway: Google is shifting from model-powered products to a unified agent harness (Antigravity) that powers agentic experiences across its entire product suite, making agents the new through-line. Google DeepMind's Logan Kilpatrick, who leads Google AI Studio and the Gemini API, explains how this agent layer connects everything from coding tools to consumer apps like Gemini and Search. He notes that while early fears of cannibalization exist, agentic AI is proving positive-sum, driving more usage and outcomes rather than reducing human engagement. A memorable insight: 'Success for Google probably doesn't look like maximizing eyeball time in front of our products. It's like maximizing outcome for customers.' Kilpatrick highlights Omni as a unified world model for generative media that excels at subtle video editing, and discusses vibecoding's rapid progress, with hundreds of thousands of Android apps already built in AI Studio. He emphasizes that coding agents are accelerating internal development dramatically and predicts vertical superintelligences emerging first in verifiable domains like science and finance.

Replit CEO Amjad Masad shared that the Replit Agent team made Fable remarkably affordable and reliable, enabling a complete state of flow with zero frustration. He demonstrated building an entire company canvas including web app, mobile app, marketing, and App Store materials all in one place. This marks a shift where cheaper and faster models may soon suffice for most vibecoding needs.

Swyx congratulated friends at Ona HQ on joining OpenAI, sharing a talk with alpha on what's next for Codex. Thibault Sottiaux from OpenAI expressed excitement about working with Johannes and team. Sam Altman also welcomed the collaboration. This move signals OpenAI's push to enhance its coding capabilities.

Vercel CEO Guillermo Rauch highlighted the powerful Vercel + Shopify combo, where a custom Next.js storefront processed 500+ orders in 2 minutes, built with v0 and Cursor. He also announced Vercel + Grok integration. Peter Steinberger shared using Codex for PRs and moving media conversion to WASM in OpenClaw for better security.

Anthropic Engineering: How we contain Claude across products. The blog explains shifting to environment-based containment like ephemeral containers, human-in-the-loop sandboxes, and sealed VMs to cap blast radius as agent capabilities grow. Key lessons include approval fatigue issues, pre-trust prompt risks, and the need for defensive proxies. Despite model improvements, deterministic boundaries remain essential. Link: https://www.anthropic.com/engineering/how-we-contain-claude

Swyx argued that stacking loops effectively—knowing when to go down for reliability or up for leverage—is key for the next century. He is building his own vibecoding platform to close the loop on errors and reduce fragmented infra setup across tools like PostHog and Arize.

Box CEO Aaron Levie shared a survey of 1,640 IT leaders showing companies adopting AI most aggressively are planning the largest headcount growth, as productivity gains lead to more ambitious projects and reinvestment rather than job cuts.

Y Combinator CEO Garry Tan critiqued a piece questioning gifted programs, noting the 12.3% eminence rate in a gifted cohort far exceeds general population base rates, calling out flawed understanding.

Every CEO Dan Shipper tried a big project on Fable but hit safeguards and switched back to Codex, highlighting current limitations in long-running agent runs.