AI Builders
Digest

Anthropic Research's Alex Albert highlights favorite features in the new Claude Opus 4.7, including excellent async work and instruction following, more predictable effort levels with a new xhigh option, no downscaling of high-res images, and noticeably better taste in UIs, slides, and docs. The official Claude account announces the model is now available on claude.ai, the Platform, and major clouds, with Claude Code updates like the new /ultrareview command for dedicated code reviews and extended auto mode for Max users. Anthropic's Cat Wu and Thariq share practical tips such as setting default effort to xhigh, adding verification workflows to claude.md, and new docs sections for what's new and webinars. These updates make agentic coding more reliable and efficient for developers.

OpenAI VP Science Kevin Weil and CEO Sam Altman announce major improvements to Codex, now with shockingly good computer use capabilities, an in-app browser, many new plugins, and the ability to learn from experience and proactively suggest actions. Codex can now use all apps on your Mac in parallel without interfering with your work. Additionally, they launched GPT-Rosalind, a frontier model specialized for biology, drug discovery, and translational medicine with built-in knowledge of relevant databases and tools, plus a Life Sciences plugin for Codex available to everyone. Box CEO Aaron Levie notes this enables agents to execute long-running tasks like drafting reports, reviewing contracts, and processing enterprise content using the Box plugin.

Anthropic Engineering explains how Managed Agents decouples the 'brain' (Claude and harness) from the 'hands' (sandboxes and tools) and persistent session log, virtualizing agent components for durability, scalability, and future-proofing. This solves previous pet-container issues, enables VPC integrations without peering, drastically reduces time-to-first-token, and supports many brains and hands. The Claude Blog highlights how it lets teams ship production agents 10x faster with secure sandboxing, long-running sessions, multi-agent coordination, and trusted governance. Early adopters like Notion, Rakuten, Asana, Vibecode, and Sentry are already using it for coding agents, productivity agents, finance/legal workflows, and bug fixing.

Notion is dedicated to being the best system of record for enterprise work. In the Latent Space podcast, Simon Last and Sarah Sachs from Notion discuss their four or five rebuilds of custom agents since late 2022, revealing that early efforts failed due to immature models lacking function calling and sufficient context. The big unlock came with advanced reasoning models, enabling reliable background agents with sophisticated permissioning. Simon is bullish on CLIs for powerful self-debugging coding agents in terminal environments, while still seeing value in MCPs for narrow, tightly permissioned lightweight agents. They are excited about the 'software factory' where teams of agents collaboratively develop, debug, review, and maintain codebases. Key leadership lessons include a low-ego culture comfortable deleting code, focusing on user journeys like email triaging rather than cool tools, and distributing tool ownership through evals. 'MCP is just the dumb simple thing that works, and it is pretty good,' notes Simon. The custom agents launch was Notion's most successful yet, boosted by free trials.

Replit CEO Amjad Masad highlights how you can turn your web app into an iOS app for less than $10 using their tools. He also announced 50% off on features especially useful for running parallel agents to accelerate project progress, and new Deploy to EU capabilities for better global performance.

Vercel CEO Guillermo Rauch introduces Workflow SDK as the solution for the hardest problem in agents and backends: durability. It handles LLM downtime, rate limits, and database slowdowns with the reliability of SQS/Kafka/microservices but without the complexity, featuring self-hosting and multi-cloud support from day one. He also praises the Vercel AI SDK and AI Gateway for capturing the upside of AI progress volatility across models like the new Anthropic ship.

Cursor_ai Design Lead Ryo Lu has been working on Baby Glass with Flowstated, a new prototyping environment that lets designers quickly imagine, remix, and test ideas in code using the same shared components powering Cursor 3’s new interface. He also shares his personal workflow using Opus 4.7 for planning, Composer 2 for building and iterations, and Codex/GPT-5.4 for hard bugs—all within Cursor_ai.

Anthropic's Cat Wu shares Claude Code tips: tell the model how to verify changes by putting testing workflows in claude.md or adding /verify-app skill, and they've set default effort level for Opus 4.7 to xhigh. Thariq announces updates including a curated What’s New section in docs, monthly 'what we shipped' webinars, and more on /usage to help users maximize Claude Code.

Product at Roblox Peter Yang recommends always asking AI to spin up a separate eval agent for yes/no checks on the first agent's output, looping until all checks pass—currently building one for YouTube thumbnails and titles. He also suggests Claude and Codex apps should support easy viewing and editing of markdown files for hand-crafting final instructions, and humorously notes the cycle of migrating between tools like openclaw, hermes, perplexity computer, and back.

AI builder Swyx notes that scheduling slop cannons like @_lopopolo on day 1 and grown ups like @badlogicgames on day 2 at @aiDotEngineer perfectly represents the most important split in AI engineering right now. He shares links to Slow Down and Harness Engineering talks.

Y Combinator President & CEO Garry Tan shares bug fixes for the /ship skill and major GBrain security fixes. He also built proper e2e Gemini Live tests for GBrain Voice with his claw, with open source release coming soon.

OpenClaw's Peter Steinberger celebrates almost 2000 people at ClawCon Michigan, calling it a weird lobster cult. He also reflects that OpenClaw's many GHSAs are actually an indicator of the coming storm in security for AI agents.

Builder Zara Zhang shares her top three time investments: deep talks, deep reads, and deep play with AI, while turning frontend slides into HTML videos declaring 'HTML is eating everything.' Every CEO Dan Shipper ran a philosopher draft on which historical philosopher each model lab would hire. Aditya Agarwal jokes about rewriting all code with 4.7. Linear Head of Product Nan Yu observes the AI industry rapidly reinventing financial concepts like crypto did, but faster, including for engineers and PMs.