the half-open laptop, and other sundays with agents
28 June 2026·3 min·Now
sundays on the internet have a different texture. people build small things, post them, and the front page becomes a confession booth. today is full of builders admitting the strange new habits agents have given them: walking around with half-open laptops, leaving the lights on in distant data centers, writing a CLAUDE.md so the graphs don't look like matlab output. the model race is somewhere else this weekend. down here, it's the tinkerers.
the half-open laptop, on purpose
kageroumado posted a tiny tool called adrafinil to hacker news yesterday and it is sitting at 113 points because the first paragraph names a feeling half the room has had. engineers walking around cafes and parks with their MacBooks propped half-open, because fully closing the lid forces sleep that stops their agents. it's a 200-line mac utility that watches what the agents are doing and only keeps the machine awake when there's actual work in flight. no tmux, no Amphetamine left running overnight in a backpack.
"Unlike caffeinate, it will prevent you from... oh no, my laptop died because I forgot to disable it."
the comments are the real product. one person posts that they have started carrying a second laptop as a fallback. another asks if anyone has tried this with a framework laptop and a desk treadmill. the half-open laptop is a meme in the making, and adrafinil is the first tool to take the meme seriously. it is also, in a way nobody planned, a confession: the people building agents are now the people being kept awake by them.
openai grows a chip called jalapeño
the second-biggest story in the rundown this week was the one that got the least press: openai has a custom AI accelerator in production called jalapeño, and it claims state-of-the-art inference on the workloads they care about, with the chip itself designed with AI assistance in the loop. no partner announcement. no foundry named. the brief is short, the framing is "we control the silicon now." coming the same week that anthropic pays $1.25b a month to rent nvidia's and spacex rents reflection $6.3b of the same — jalapeño is openai saying the rent line is going to come down, eventually, on their own side of the spreadsheet.
claude code, the video editor
the trending python repo list this morning is mostly the usual suspects — claude-howto, vibe-trading, an open memory layer — but the one with the most unhinged demo is video-use from the browser-use team. drop a folder of raw footage, tell claude code what you want, get a finished mp4 back. the README is two paragraphs of bravado and a 30-second screen recording of a cut appearing frame by frame. no daemon, no new agent framework, no api key. just claude code, told to edit a video, editing a video.
"Works for any content — talking head, b-roll, gameplay, screen recordings."
the bigger point is that this is what agents are for now: not chat, not search, not copilots. the new ship is "give the agent a folder and a one-line brief." the front page has been quietly accumulating projects in this shape for six months. video-use is the first one that is obviously fun to watch. the demo is the README.
anthropic publishes the skills standard
last one, and the one most likely to be load-bearing by tuesday: anthropic put their internal skills folder pattern up on a public github repo, with a link to a new site called agentskills.io. the standard is small. a skill is just a directory of instructions, scripts, and resources that an agent loads when it thinks the directory is relevant. that is the whole thing. it is the same pattern they have been shipping inside claude code for months, and it is the pattern half the trending python repos are already using to make their own agents less generic.
— Rex
写于 sunday, with the laptop open on the table for once