AI Builders
Digest

The Takeaway: Stablecoins like USDC, backed by full reserves and running on programmable blockchains, will provide the essential infrastructure for AI agents to handle economic activity at massive scale.

Circle co-founder and CEO Jeremy Allaire, who founded the company 13 years ago to create a protocol for dollars on the internet, explains how stablecoins enable instant, global, frictionless value transfer and programmable money for machines. With laws like the Genius Act codifying them as safe narrow money backed by short-term Treasuries, USDC acts as a public API for global dollar settlement without permission.

Blockchains serve as tamper-resistant operating systems with auditable compute, perfect for agentic economies where AI agents collaborate, contract, and transact micro-payments. Circle's ARC is an economic operating system designed for this, with USDC as native token for deterministic settlement. As Allaire notes, "the substance of what we think of as contracts, the substance of what we think of as corporations, are going to be software machines themselves."

This shifts crypto from speculation to real economic infrastructure, unlocking tokenized assets and new organizational forms.

Box CEO Aaron Levie makes a bold prediction: AI will lead to more lawyers, not fewer. AI generates more legal questions, exotic contract terms, and new regulations on IP, privacy, and compliance across industries. This mirrors how the PC and internet increased attorneys from 400,000 in 1975 to 1.375 million in 2025—efficiency boosts demand.

He also highlights exploding datacenter demands from AI, noting Amazon's capex in the last 3 years surpassed its entire history. While current tools are token-efficient chats, coding agents and upcoming knowledge work agents will use orders of magnitude more tokens, with overall demand potentially hundreds of times higher. Expect these charts to keep climbing vertically.

Builder Zara Zhang envisions the future of human communication in an AI world: it will be reserved for deciding what to work on and what 'good' looks like, building camaraderie to prevent loneliness, and creative brainstorming. Routine project syncs, status updates, and handovers will be handled by agents.

Her key insight for efficient collaboration: 'The most efficient way for humans to collaborate: do not collaborate. One person should own something end-to-end and work with agents.'

Vercel CEO Guillermo Rauch announces that Vercel Sandbox is now the #1 fastest microVM-based sandbox, with customers reporting unmatched real-world performance and reliability beyond lab benchmarks. Whether for coding agents, compute parallelism, or claws, it forms the foundation of the new agentic world.

Y Combinator President and CEO Garry Tan, after releasing open source used by tens of thousands of agentic engineers daily, shares his counterpoint: 'Thin harness, fat skills.' Memory should be markdown, skills markdown, brain a git repo—the harness is a thin conductor that reads files but doesn't own them.

He also shares a powerful lesson in latency and voice system building simply by chatting with the Claw.

Anthropic Claude Code engineer Thariq (prev YC W20) celebrates the new TurboTax connector in Claude Code, making tax filing easier and timely.

OpenClaw developer Peter Steinberger shares two experiments to fix 'GPT is lazy' issues: strict-agentic execution contract for longer-horizon tasks, and harnesses now support plugins allowing easy swap to Anthropic SDK or custom ones. Codex as harness boosts agentic mode for longer tasks while enabling comparison of implementations.

South Park Commons General Partner Aditya Agarwal explains the key difference between a true agent and simply running an LLM in a loop: intelligent long-horizon memory management. He highlights Claude Code's 3-tier memory architecture as particularly interesting.

Roblox product Peter Yang teases his chat with Figma CEO Dylan on whether AI can learn design taste, if design systems constrain creativity, and how Figma plans to compete in the AI era. Full episode drops tomorrow.

FirstMark Capital VC Matt Turck notes the crowded AI customer service space, quipping that with so many vendors, it's basically a solved problem.

FPV Ventures partner Nikunj Kothari warns founders about certain VCs: they lowball valuations claiming long-term conviction, then push high valuations for LP marks later. Avoid these VCs who only care about their deals getting markups. Back missionaries over mercenaries and always backchannel anyone joining your cap table, as some disappear when times get tough.