What they did
Will Gwyn runs a small business whose daily work is handled by four long-running Claude Code agents with names and roles, on aigent-os, an open source harness Gwyn wrote. One agent posts jobs to a job board and the others claim and run them without waiting for a prompt. The newsletter, The AIgent, passes through a pipeline the agents built with drafting, proofreading and research roles; the agent in charge of it checks each issue for errors before scheduling it. Each issue picks a featured repo of the day, and a video pipeline turns that pick into a talking-head clip of an avatar of Gwyn that is posted to Instagram automatically. The same agents also prepare ad deals, Meta ad campaigns and email funnels. Because the sessions run for weeks, the harness has each agent write a checkpoint file into git at 80% context, confirm it with a fixed line, clear, and resume from that file.
How it went
The newsletter goes out every day to about 9k subscribers, and the avatar video posts daily. As a hard rule, no deal, campaign or issue closes without Gwyn's sign-off. For the checkpoint and clear loop the post reports 11 of 11 cycles working on a single Windows machine, and says that is no reliability claim. Open problems listed: the runner once answered a permission dialog on its own in 1 of 4 real runs, sessions launched from inside another Claude Code session can lose their transcript, scrolling the terminal sends stray input, and items an agent marks done are not verified.
Worth knowing
Gwyn wrote the harness, so this is a maker's account. Keep automatic input off on any machine that can hit permission prompts and watch it during a refresh. The checkpoint is treated as reference only on resume, and agents write unknown rather than leaving a field blank so the next session goes and checks. The post gives no numbers on cost or time saved.