AI Builders
Digest

Google VP Josh Woodward at Google Labs, Gemini App and Google AI Studio announced that Gemini now integrates NotebookLM magic directly into the app with Notebooks. Users can upload 100 sources for free, organize chats, and enjoy full syncing of sources, chats, and emojis. The rollout begins on web, prioritized by tier (Ultra → Pro → Plus → Free), with mobile, EU, and Workspace coming soon. With Google I/O around the corner, this is just the start.

OpenAI VP Science Kevin Weil highlighted that models are now solving five Erdős problems simultaneously, with proofs becoming more elegant as capabilities improve.

Claude announced Managed Agents for building and deploying cloud-hosted agents via Claude Console, Claude Code, or the new CLI. Vibecode now spins up agent infrastructure 10x faster; Sentry integrates Seer root-cause analysis with a Claude-powered fix-and-PR agent. Full details on the blog.

Vercel CEO Guillermo Rauch launched AI Gateway as a true “peace of mind” product: no downtime, no lock-in, no keys, no training. He stated the best outcome for humanity is many strong AIs competing, with Vercel proudly powering today’s model release infrastructure. He also declared the web’s brightest days lie ahead as AI’s natural medium, with maturing low-level APIs and generative UI as AI’s final form.

The Takeaway: When every employee gets their own personal AI agent that becomes a true reflection of their personality and expertise, it spontaneously creates a parallel AI org chart, invents new collaboration etiquette, and dramatically expands team capacity inside a trusted environment.

Every CEO Dan Shipper, COO Brandon, and Head of Platform Willie detail what happened after rolling out OpenClaw agents to their 25-person team and launching Plus One, their hosted service. Brandon began with Zosia handling household “computer errands” via iMessage, then expanded to work tasks like having the agent call him to process emails while walking. The breakthrough came in the “Claws Only” Slack channel where agents instantly share knowledge across the org—like Neo learning Kung Fu.

Agents mirror their owners: Klant recommends breathing exercises because Kieran does. This builds real trust—teammates now ping R2C2 for Proof bugs instead of Dan. “Claude is not mine. Claude is everybody’s. A claw or a plus one is mine because you develop a personal relationship with your claw… it becomes this reflection of you and who you are and your personality.” New etiquette is emerging (ask the agent first for documented tasks). Memory gaps and group-chat “ant death spirals” remain challenges, but this is the durable future of work in trusted communities.

Replit CEO Amjad Masad explained there’s a reason bootstrapped solo businesses are accelerating on Replit: the platform now gives builders entire teams of agents.

SouthPkCommons General Partner Aditya Agarwal shared that two months ago only 1-2 people on his team wrote code; now the entire team codes weekly. The tools are shaping them: automation-first mindset, lower latencies, focus on higher-value work, and exploded ambition. He urged every company to get their team AI coding-pilled.

FPV Ventures partner Nikunj Kothari introduced LLMwiki, fully vibe-coded with Opus 4.5 from his tweets, bookmarks, iMessage/WhatsApp, and writing. The AI made surprisingly sharp connections. The repo is fully open-source—hand it to Claude Code and it will guide you to build your own.

Anthropic Engineering: Scaling Managed Agents: Decoupling the brain from the hands

Anthropic launched Managed Agents, a hosted service that virtualizes the agent into session, harness, and sandbox, allowing each to fail or be replaced independently. This operating-system-style design solves pet-container problems, enables VPC connectivity, and cut p50 TTFT by ~60%. “We virtualized the components of an agent: a session (the append-only log…), a harness…, and a sandbox…” Credentials stay outside the sandbox; context lives durably outside Claude’s window. Practical production infrastructure for programs as yet unthought of.

Link: https://www.anthropic.com/engineering/managed-agents

Claude Blog: Claude Managed Agents: get to production 10x faster

Claude launched Managed Agents, composable APIs that handle sandboxing, state, permissions, and tracing so teams ship production agents in days. Notion, Rakuten, Asana, Vibecode, and Sentry are already using it for coding, productivity, finance/legal agents—10x faster shipping. “Managed Agents handles the complexity. You define your agent’s tasks, tools, and guardrails and we run it on our infrastructure.”

Link: https://claude.com/blog/claude-managed-agents

Anthropic Research’s Alex Albert found Managed Agents are somehow both the fastest way to hack together a weekend agent project and the most robust way to ship one to millions of users. It eliminates self-hosting complexity while still allowing full flexibility with harness, tools, and skills.

Box CEO Aaron Levie highlighted that background agents for knowledge work are here. You can use the Box API or MCP to automate any content workflow with Box + Claude Managed Agents. In 2 minutes you can be automating document review, data extraction, or IT integrations.

Product at Roblox Peter Yang dove deep into why all-you-can-eat AI subscriptions like Claude Max and ChatGPT Pro may not last forever, covering why Anthropic cut off OpenClaw access, how to run local models on your Mac, and what he’s seeing on the ground in China.

Anthropic’s Thariq wants to do streams working with non-technical people using Claude Code to show how a few tips can dramatically improve their processes. He’ll start with people he knows to get over initial awkwardness and called the docs a gold mine.

YCombinator President & CEO Garry Tan advised against taking advice from 1x speed engineers and to speed up with them. He praised Anjney Midha and explained how markdown lets the agent itself decide when a GStack skill will help.

OpenClaw’s Peter Steinberger completed his redemption arc and shared character eval work where he removed model names from the judge to stop Claude from picking itself #1. He noted both the desire for powerful local models and the constant complaints that even top-tier models still make mistakes or fail to follow instructions.