← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-04-12

2026-04-12 · Historical edition

X / Twitter

Box CEO Aaron Levie

Levie produced the three most information-rich posts in this batch this week. First, sharing a data-center capital expenditure chart, he noted that Amazon's capex over the past three years exceeded all its prior spending combined. AI applications still largely consist of relatively low-token chat tools; widespread adoption of token-intensive uses such as coding agents would increase compute demand by orders of magnitude. Second, he invoked the Jevons paradox in security: widespread AI tools will increase demand for security professionals, because automated exploitation provides scanning at scale, while evaluating and fixing vulnerabilities still requires human expertise. Finally, he reported findings from speaking with dozens of IT and AI leaders across banking, media, retail, healthcare, consulting, technology, and sports: enterprise AI is clearly moving from the chat era to the agent era. That post received more than 2,900 likes.

OpenClaw ClawFather Peter Steinberger

Steinberger announced two experimental features for OpenClaw's next release. Strict Mode uses `executionContract: "strict-agentic"` to force GPT-5.x to keep working instead of stopping prematurely. The second uses Codex as OpenClaw's underlying harness, letting it handle thread management, resuming interrupted runs, context compaction, and app-server execution. He also announced that harnesses can now be swapped as plugins, making comparisons with the Anthropic SDK and other implementations easier. The posts collectively received more than 1,300 likes, reflecting strong developer interest.

Builder Zara Zhang

Zhang shared a series of concise, substantive observations this week. She argued that the most efficient human collaboration is “not collaborating”: one person owns something end to end and works with agents. She predicts that human communication will retain only three functions: deciding what to do and what counts as good, providing companionship to prevent loneliness, and creative brainstorming. Sync meetings, status reports, and handoffs would all disappear. She also vibe coded a personal Chrome new-tab page, with features including grouped tabs across domains and sound effects and confetti when closing tabs. She open-sourced the code, and the post received more than 940 likes.

Vercel CEO Guillermo Rauch

Rauch shared a customer relationship practice: create an X group chat for v0's most demanding core users and engineering leads, ask members for high-quality feedback, and have engineers respond in real time and ship quickly. He described this as “face the music”—facing users directly—and invited people to DM him to join. The post received more than 540 likes.

Replit CEO Amjad Masad

On Apple's 50th anniversary, Masad sharply criticized the company, implying that it was working toward becoming “the most hated company.” Against the backdrop of Apple's recent disputes with developers and regulators, the post received more than 8,000 likes.

Y Combinator CEO Garry Tan

Tan shared a central view on agent architecture: “If your memory disappears when the harness crashes, your harness is too thick. Memory is Markdown, Skills are Markdown, the brain is a Git repo, and the harness is just a thin orchestration layer—it reads files, it doesn't own them.” He defined “thin harness, fat skills” as an architectural principle for agentic engineers. The post received more than 100 likes.

Anthropic Claude Code Engineer Thariq

Thariq humorously announced that Claude Code now supports a TurboTax connector, joking that he could finally procrastinate on filing his taxes openly. The post received more than 1,400 likes, reflecting developers' surprise at the breadth of Claude Code's expanding ecosystem.

South Park Commons Partner Aditya Agarwal

Agarwal said long-horizon memory management is the key distinction between a real agent and “running an LLM in a loop.” He praised Claude Code's three-tier memory architecture as one of the most worthwhile examples of agent engineering to study so far.

FPV Ventures Partner Nikunj Kothari

Nikunj shared a startup growth insight he calls the “DoorDash path”: DoorDash beat Uber Eats precisely by focusing on suburban markets outside cities, rather than competing head-on with major players in core markets. He believes more founders should pursue this approach of breaking through by focusing on “secondary markets.”

Peter Yang (Roblox Product)

Yang raised a widely discussed question: with many community members reporting that Claude Opus had been “nerfed,” he asked, “Why would Anthropic deliberately weaken its own model?” The post received more than 530 likes, showing that model reliability has become a major concern for developers.