There's a quiet theme running through today's items: what happens to your judgment when AI does the work for you. Vercel CEO Guillermo Rauch named it with one concept. A video AI team learned it the hard way by trying to give agents eyes they don't have. Worth sitting with both before your next sprint planning.
01
Coding agents will make you love your own bad decisions
Vercel CEO Guillermo Rauch posted a one-liner that's been rattling around since Sunday: "Coding agents will squeeze every ounce of IKEA effect out of you, if you let them." The IKEA effect is the psychological bias where you overvalue something because you helped build it. The problem with agents is that you barely did build it, but you still feel like you did, and that feeling will make you defend mediocre output you'd reject from a contractor. ---
Why it matters: If you're reviewing AI-generated code and thinking "good enough," ask whether you're making a technical judgment or just rationalizing something you feel ownership over. Your code review process needs to account for the fact that agents make you a worse critic of their work, not a better one.
AI video agents can't see. So this team taught them to write HTML instead.
Peter Yang shared a lesson from a builder named Liu Bin who tried to build a video agent and discovered the hard limit fast: AI agents have no real visual intelligence. They can't look at a frame and understand whether it works. The team's fix was to stop treating video as a visual medium and start treating it as code. HTML, CSS, and JavaScript became the production layer, with footage and images sitting on top. The agent writes what the video should do; the browser renders it. ---
Why it matters: This is a useful mental model for anyone building with multimodal AI right now. When an agent fails at a visual task, the instinct is to find a better vision model. The better move might be to reroute around vision entirely and find a representation the model is actually good at. Code beats pixels.
Codex usage banking is revealing how developers think about scarcity
Thibault Sottiaux, who works on the Codex App at OpenAI, asked users a behavioral question after the platform added the ability to bank unused session resets: are you hoarding them, or burning through without hesitation? The post got 868 replies, which is a lot for a product question that sounds simple. ---
Why it matters: The way developers ration AI usage tells you something real about whether they trust the tool. If people are hoarding resets, they're treating Codex like a limited resource to be spent carefully, not an always-on assistant. That psychology shapes how much of the tool's potential actually gets used. OpenAI has a product design problem if power users feel like they need to save up.
"Quality is irrational" — Nan Yu on what it takes to control your own stack
Nan Yu posted a reflection on a framing he found useful: choosing quality consistently requires "an irrational level of commitment" and "an irrational level of self-belief that controlling things top to bottom will yield better results than using common frameworks." This connects to yesterday's Rauch post in a direct way. If agents erode your attachment to craft, and frameworks erode your control over the stack, you end up with code nobody really owns. ---
Swyx on an AI insurance startup breaking through in a frozen market
Shawn Wang (known as Swyx), an independent AI researcher and builder, noted while shopping insurance for his new media lab that Corgi, an AI-native insurer, has apparently achieved something rare: his real estate broker told him Corgi is covering "every single one of my clients right now." Swyx called it "~100% greenfield market share" in a sector that almost never moves this fast. Worth a mental note: insurance is one of those industries where AI-native distribution is supposed to be theoretically possible but practically slow. If Corgi is actually moving this fast via referrals, that's a different story.