The AI agent is now doing your IT support tickets. Installing browser extensions. Filing GitHub issues. Fetching your Shire renders. And Thibault Sottiaux at OpenAI just quietly revealed that even AI users take the weekend off. The picture emerging this week is less "robots taking over" and more "robots doing the annoying small stuff, while humans spend Friday night relaxing."
01
Andrej Karpathy gave Claude Opus 5 a $10 budget and two hours to render Lord of the Rings
The former Tesla AI director handed the first paragraph of Fellowship of the Ring to Opus 5 with a 1-million-token budget and the instruction to render it in three.js. Two hours and 5,500 lines of procedurally generated code later, the model had built an interactive visual scene. Karpathy's framing: we're leaving the era where you test an AI by asking it to draw a pelican on a bicycle. The new tests are open-ended creative engineering tasks that run for hours and cost real money. ---
Why it matters: The "$10 prompt" is the new benchmark. If your team is still evaluating AI models by asking them to write a function or summarize a PDF, you're testing the 2024 version of these tools. The ceiling has moved. The question worth asking internally is whether your workflows have moved with it.
Vercel's Guillermo Rauch spotlights an open-source agentic CRM
Rauch highlighted a new open-source CRM built on Vercel and Next.js. It's model-agnostic, self-hostable or serverlessly deployed, multi-channel, and headless, meaning it has no fixed front-end and connects to other tools. He called it "the way," which from Rauch is a genuine endorsement, not a throwaway. ---
Why it matters: Yesterday we covered Rauch describing the Issue > Agent > PR > Release loop. An agentic CRM is that same loop applied to sales and customer ops. If your sales team is still copy-pasting email threads into Salesforce by hand, this category of tool is coming for that workflow faster than most CRM vendors want to admit.
Nan Yu wants GitHub to let you pay AI agents to fix your bugs
The Linear product lead proposed a feature: when you open an issue in an open-source repo, you should be able to pledge compute tokens alongside a written spec. If the maintainer accepts, GitHub routes the issue directly to a cloud coding agent, billed to the requester. The maintainer reviews the output. No more low-effort pull requests from contributors who didn't read the codebase. ---
Why it matters: Open source maintainers spend enormous time reviewing bad PRs from people who wanted the bug fixed but didn't want to do the work. This flips the model: if you want a fix and you're willing to pay for the agent to attempt it, the maintainer gets a clean spec and a draft solution instead of a half-baked patch. GitHub would have to build it, but the economic logic is straightforward.
Peter Steinberger let his agent fix a years-old annoyance in Gmail
The developer shared a small but telling moment: he'd been bothered for years by Gmail's default appearance and never got around to fixing it. He finally asked his agent. It installed a fix for him. ---
Why it matters: This is the actual use case that spreads AI agent adoption past early adopters. Not the $10 Tolkien render. The thing you've been tolerating for three years because the fix was annoying enough to keep deprioritizing. When the activation energy drops to "just ask," the list of things people actually fix gets very long, very fast.
OpenAI's usage data shows AI users slow down on weekends
Thibault Sottiaux, who runs product for OpenAI's coding tools, shared a small data point: users invoke the `/fast` command significantly less on weekends. His read: "The weekend is for relaxation, even for the model."
Why it matters: This is a throwaway line with a non-throwaway implication. If AI usage tracks human work rhythms this closely, the "AI never sleeps" narrative companies use to justify headcount decisions is at least partially wrong. People are still driving these tools. When people rest, the tools rest too.