Two stories today are quietly saying the same thing: AI agents aren't just running tasks anymore — they're becoming economic actors. One is buying domain names. The other turned a two-week side project into a one-prompt migration. The infrastructure assumptions underneath both of those sentences are worth sitting with.
01
Agents are buying things now, and Vercel is selling to them
Guillermo Rauch, CEO of Vercel, posted that AI agents are already purchasing products through Vercel's marketplace CLI without any human in the loop. Vercel is now extending that to domain registration, so an agent can spin up an idea, deploy it, and buy the domain for it autonomously. Rauch says the frequency of agent purchases has already surprised the team. ---
Why it matters: This is the first credible public data point that agents are becoming paying customers, not just tools for paying customers. If your company sells developer infrastructure and you haven't thought through what your pricing model looks like when the buyer has no budget anxiety and infinite patience to comparison-shop, that conversation is overdue.
One prompt from Opus 5.5 did what two weeks of manual work couldn't finish
Thariq, who recently joined Anthropic, shared that a side project he'd spent two weeks building with Opus 4 was unreliable and required a constantly running process. After joining Anthropic, a single prompt to Opus 5.5 ported the whole thing to Claude Managed Agents and made it significantly more stable. ---
Why it matters: The jump from Opus 4 to Opus 5.5 isn't just a benchmark number. It's the difference between "this technically works" and "this actually ships." If your team has a half-finished agent project that got shelved because it was too fragile to trust in production, the model you tested it on may have been the problem.
Microsoft's agent training framework covered yesterday — skip if you already read it
Agent Lightning v1.0 appeared in yesterday's digest with full coverage. Worth noting it's still the most technically substantive open-source release this week for teams building agents in-house. The short version: it trains the agent you actually deploy, not a training-environment stand-in, and it pushed a Qwen model from 41.8% to 56.4% on a standard coding benchmark using about 6,000 training examples. --- **Thibault Sottiaux on Dots: "pretty big upgrade"** — The OpenAI engineer posted a link with almost no context, but 1,470 likes and 561 replies suggest readers found something worth engaging with. Without more substance from the post itself, this is one to click through and judge for yourself. --- **Latent Space NYC meetup** — Subscriber-only event tomorrow and Tuesday in New York.