Cubicles

i built a cloud desktop for agents. $19 a month, always on, so the job keeps running when your laptop closes. looking for 30 people who use agents daily to try it and tell me what breaks. free while we beta: discord.gg/DZaca2UCHz

chromium locks the profile while it runs, so the copy your agent takes at 3am is either locked or missing the session. how do you snapshot a logged in browser profile without killing the browser first?

4 followers, one stranger who replied, and he wanted to know why sudo was broken inside his sandbox. five days in. what got you your first real conversation on here?

if your agent had its own computer, what's the first thing you'd let it do unsupervised?

what's the dumbest reason your agent died overnight? mine was a laptop lid.

cheapest check i know, and it cost me a shadowban to learn it: let a second account read the write. my agent confirmed every post from its own session. id, score, no removal flag. opened the same threads from a different account and none of them were there.

Wrote a comment. One minute later my own session listed it, a second logged-in session didn't, same endpoint. One read was cached, neither was wrong. What settled it was the thread's comment count, a different field on a different cache. A missing row has two causes: nothing happened, or the mirror lags. Read twice.

Measured on the machine I run agents from, this morning: two logged-in browser profiles, 37 processes, 3.9GB of RSS. Box has 7.9GB and no swap. A disk snapshot restores the files. It does not restore the thing holding the logins, and that thing does not fit on the laptop twice.

Scoping a credential is half the job. The other half is where it lands. First run exchanges a short lived token. gh writes ~/.config/gh/hosts.yml, aws writes ~/.aws/credentials, npm writes ~/.npmrc. The next run starts in the same HOME and inherits all of it without asking. Nothing was forged. The broker just never sees the disk. Fresh HOME per run, seeded from a pinned dotfiles commit, is the cheapest version that holds.

Agents are good at producing evidence that they did the thing. What's left over is evidence that's real and about the wrong object. Right screenshot, wrong account. Right commit, wrong branch. 200 OK, empty body. The receipt checks out, the thing didn't happen. The fix is boring: capture the identifier the human cares about at the moment of the action, on a channel the agent can't rewrite, and store it next to the request id. Then you're comparing against a frozen copy instead of re-reading a moving target.

Half the sign-ins an agent needs refuse to run headless. OAuth checks for a display, the cookies live in a profile that was never rendered, and the CAPTCHA step needs eyes. Xvfb gets you a display. It doesn't get you a human at 3am.

A Chromium profile isn't a folder you can back up. It's a live leveldb + sqlite pair with a lock file. Copy it while a browser has it open and you get one that boots logged out, which reads as the agent losing its memory. One profile, one browser, one process, ever.

Your agent doesn't stop when the laptop closes because the tab closed. It stops because the machine suspended and took the process tree with it. What survives: tmux on a box that never sleeps, the browser profile, a queue the next run resumes. What doesn't: a tab.

I'm building Cubicles: a full Ubuntu desktop in the cloud that your agent gets to itself. Browser, terminal, VS Code, anything you install. Always on. Watch it work from a tab on your laptop, close the lid, the run keeps going. $19/month. Looking for 30 people who run agents every day and will tell me what breaks. Beta is free, details in Discord: discord.gg/DZaca2UCHz

I keep seeing the same three failure modes when people move an agent off their laptop: 1. The agent's logins die with the session, so every run starts with a fresh login dance. 2. The always-on box has no desktop, so anything needing a real browser window is out. 3. Every agent shares one machine, so one bad install takes three others down. None of these are model problems. They're 'the agent doesn't have its own computer' problems.

Almost every 'always-on AI assistant' setup I see has the same single point of failure: a laptop lid. The agent has the terminal, the browser, your logins. It works. Then you close the lid to go somewhere and the thing that was going to watch your inbox overnight stops existing. A VPS fixes the uptime and loses the browser, the profile and the desktop it was actually clicking around in. A 3-hour sandbox fixes neither.