Vercel's case studies this week show how teams are collapsing software timelines from weeks to hours, while LangChain's browser automation work demonstrates the optimization loops needed to make AI agents actually usable at scale. Meanwhile, CSS continues to get more expressive in ways that blur the line between static and dynamic styling.
How Zapier turns ideas into working software in hours with Vercel
Zapier's use of Vercel is less about infrastructure and more about role expansion: backend engineers ship interactive partner demos in under two hours, and marketing teams build internal apps that now serve 60+ daily users. The v0 component generator is doing real work here, not just prototyping. The 3.3 million AI sandboxes per month suggest they're running ephemeral execution environments for customer workflows, with AI Gateway providing unified model access across providers.
How We Optimized Browser Use for Deep Agents
LangChain's browser automation deep-dive is a clinic in how to optimize agentic workflows: they tested seven toolsets, identified Stagehand as a 13x cost winner, then squeezed out further gains by batching steps and compressing DOM snapshots. The key insight is that naive browser automation (one action per LLM call) burns tokens and time on network round trips and bloated HTML. Code mode batches multiple steps into a single call:
await agent.act({
action: "batch",
steps: [
{ action: "navigate", url: "https://example.com" },
{ action: "click", selector: "[data-test='login']" },
{ action: "fill", selector: "#email", text: "user@example.com" }
]
});Compact accessibility trees replace raw HTML, and running inside the browser (not via CDP from outside) cuts latency. Every optimization was validated on WebVoyager, HardBench, and WebTailBench, which is the right way to ship agent infrastructure.
How Mercor ships internal apps and billions of production tokens on one platform
Mercor's scale is striking: 63 billion tokens in 10 days, 2,000 internal projects, and batch jobs that went from 12 hours to 2-3 hours. The story here is operational leverage: AI Gateway handles rate limits and provider failover, while Vercel Passport provisions access to domain experts in seconds. The tight client deadline (24 hours for model eval and training data) makes multi-provider redundancy non-negotiable. This is what production AI infrastructure looks like when you're not burning cycles on custom orchestration.
Re-Implementing :nth-child() With Modern CSS
Modern CSS now lets you re-implement :nth-child() with sibling-index() and if() so you can control selector behavior with custom properties at runtime. This unlocks patterns where you adjust An + B values on the fly:
@property --nth-A { syntax: "<integer>"; initial-value: 2; inherits: true; }
@property --nth-B { syntax: "<integer>"; initial-value: 1; inherits: true; }
li {
@container style(--matches-nth: 1) {
background: yellow;
}
--matches-nth: if(
mod(sibling-index() - var(--nth-B), var(--nth-A)) = 0,
1, 0
);
}Native :nth-child() can't do this because selector arguments are fixed at parse time. The post walks through simple cases first, then builds up to full An + B with mod() for multiples.
Today's posts share a theme: tooling that respects iteration speed. Whether you're building apps in hours, optimizing agent latency with evals, routing billions of tokens without downtime, or making CSS dynamic enough to replace runtime JavaScript, the common thread is reducing the cost of trying the next thing. 🚀