A Japanese Threads creator split his account's operation into six AI agents with one deliverable each (researcher, writer, poster, fetcher, analyst, supervisor) and reports 1,181+ posts and over 10 million cumulative impressions. The writer agent may only cite numbers stored in a single 'truth file' so it cannot inflate results, the poster must confirm the published URL rather than the button click, and the fetcher records 'not retrieved' instead of zero when a fetch fails.
Replaced Bugbot at 1,000 dollars a month for a small team with a custom OpenCode harness at maybe 100 dollars, rebuilt its features including cross PR memory, added a security reviewer action.
He runs a small home service business. On site his crew capture photos and notes into a CRM; the AI reads the CRM over a connector, studies the photos and writes the full client report he used to write himself. When the report is done he tells the AI to send the invoice, and it builds the Stripe invoice from the project notes and sends it to the customer.
Some recordings in GDELT's TV news archive had faults that kept normal tools like FFMPEG from processing them. GDELT used AI 3.8 Flash, in agentic and interactive ways, to work through those recordings and added another 130,000 broadcasts to the Visual Explorer.
His AI runs four standing jobs: each evening it checks new job posts in his niche and emails him only the ones worth his time; his messy end of day notes come back as tasks with due dates in his to do app; receipts he forwards get logged into a spreadsheet by amount, date and category; and on Sundays it tells him what slipped and what he keeps avoiding.
He is letting the AI run his websites, with him watching over it. Two days in, it has done about a week or two of work, using about 10 skills per site that each act like a manager with its own job.
This only partly worked. Worth knowing before you try it.
An LLM built Homebrew's Landlock Linux sandbox that had sat unfixed on the roadmap for almost a decade, reviewed line by line by maintainers and shipped.
He gave the AI 45 podcast files, each about an hour and a half long. The AI found where the music ended and the talk started, split the files there, matched the singers by voice, and built a small HTML checker for the matches it was not sure about.