|
AI Builders Digest
Wednesday, September 16, 2026
|
|
Two stories in today's payload are essentially asking the same question from different angles: how do you keep AI agents from going off the rails? Box CEO Aaron Levie is thinking about it at the macro level (more agents, more tasks, more exposure). Vercel CEO Guillermo Rauch is thinking about it at the code level (give agents better guardrails or they'll break your design system). The answers rhyme more than it might look.
|
|
---
|
|
01
|
Guillermo Rauch: your AI agent is only as smart as the rules you give it
|
|
|
Vercel CEO Guillermo Rauch made a point that cuts through a lot of agent hype: the output quality of any AI agent is bounded by the quality of the verification tools around it. Compilers, linters, type systems, proof-checkers. He's pointing to shadcn/lint specifically as a tool that keeps agents aligned with your design system's rules. His broader framing: "verifiers plus skills are the new frameworks."
|
Why it matters: If you're shipping a product with AI agents writing or modifying code, the agent's raw capability matters less than whether it has something to check itself against. Teams without those guardrails aren't getting better output from a smarter model. They're getting faster drift.
|
|
Source →
|
|
---
|
|
02
|
Box CEO Aaron Levie says we're underestimating the agentic wave
|
|
|
Aaron Levie posted a longer-form take arguing that most people's mental model of "agentic workloads" is already obsolete. Agent swarms, better computer use, new API layers, MCP integrations, form factors like Muse and Instinct, vertical enterprise agents, background workflow agents. His point: we're going to delegate vastly more professional and personal tasks to agents than anyone initially assumed, and the information those agents will need access to is going to be far broader than what we've scoped so far.
|
Why it matters: If Levie is right, the compliance, security, and data governance decisions your company is making right now about agent access are being made at the wrong scale. The question isn't "can the agent read this folder?" It's "what happens when the agent has read everything?"
|
|
Source →
|
|
---
|
|
03
|
Anthropic opens Claude to community mods. First use case: Tetris.
|
|
|
Boris Cherny flagged that Claude Mods are now rolling out, and the community wasted no time. Someone built a playable Tetris game inside Claude within hours of launch. Cherny, who works at Anthropic on Claude Code, shared the community update with technical details and demos.
|
Why it matters: The Tetris demo is silly, but the infrastructure under it is serious. Mods mean third-party developers can now extend Claude's behavior inside the interface, not just through the API. That's the difference between Claude as a model and Claude as a platform. Anthropic is building a developer ecosystem, which changes the competitive math with ChatGPT.
|
|
Source →
|
|
---
|
|
04
|
Google's Gemini power user group is expanding
|
|
|
Josh Woodward, who leads Gemini at Google, announced that a new cohort is getting early access to upcoming features for Daily Brief and Personal Intelligence, two of Gemini's more consumer-facing capabilities. The program has been running for two months and has already tested more than 20 features.
|
Why it matters: Google is building a feedback loop between its most engaged Gemini users and its product roadmap. That's a direct counter to the criticism that Google ships features nobody asked for. Whether it changes the output is still an open question, but the structure is smarter than it was a year ago.
|
|
Source →
|
|
---
|
|
05
|
The pacing debate gets a proper dissection
|
|
|
ChinaTalk's Jordan Schneider convened Nathan Lambert (who writes Interconnects) and Jasmine Sun to work through what's actually happened since Dario Amodei's letter asking the industry to slow down. The conversation covers why the Hugging Face hack landed culturally in a way that years of extinction-risk arguments didn't, what "pacing" actually means in practice (Nathan's view: nobody can measure the frontier, but pacing still beats "pause"), who controls evaluation access (METR's monopoly on trust, CAISI's $10 million), and whether China slows down at all given Zhipu and Moonshot's continued pace. The Sacks product-liability clap-back gets a full airing.
|
Why it matters: Yesterday's digest covered Aaron Levie's take that pacing isn't regulatory capture. This piece goes deeper into the mechanics. If you work in health AI, fintech, or anything defense-adjacent, the question of who gets to evaluate frontier models before deployment is about to become your problem directly. METR having a near-monopoly on that trust layer is a structural fact worth understanding before regulators make it a structural requirement.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|