A small agency owner whose team does websites, content, Google ads and SEO for home service businesses runs the work through about 30 AI agents. The problem was memory: every session started cold and repeated mistakes. Since June they share one log that only grows, now about 11,400 lines including 413 decisions, where every note names the agent that wrote it and links to proof, and replaced facts stop showing in searches. A separate grader agent scores every piece of work, though he found 208 of 277 grades had been credited to the grader instead of the worker.
After a short session troubleshooting an iOS widget, about $1.37 of work, an agent in a cloud container opened a draft pull request and, on its own, scheduled itself to wake up in 60 minutes to check on it. Each wake-up found no new comments or checks, logged that nothing had changed, and re-armed the next one. It repeated 89 times across 9.5 days with nobody present. Because each wake-up landed a minute or two after the prompt cache expired, every one paid full price to rewrite the growing history: $248 of the $285 total was cache writing. His entire $250 credit balance drained to zero.
$285.12at stake
This one went wrong. See what happened before you try it.
He runs Qood News. When the AI 2.5 was retired, moving to the AI 3.1 would have tripled his AI bill, so a decision model called Jev now decides what counts as good news and costs stay about the same. In testing Jev approved 39 of 110 stories while the AI approved 53, mostly dropping filler.
Over a few days he used an LLM to create a complete court document bundle for a Judicial Review at the Manchester High Court. He filed the documents today and the court issued them within 5.5 hours, something he says he had no chance of doing on his own.
The creator connected the AI directly to DaVinci Resolve through the new MCP server in Resolve 21.1 and had it edit a real video by talking to it. The AI created the project, imported footage, read the transcript to find stutters and retakes, built a clean best takes cut, stripped dead air, and synced a phone screen recording to the main camera by audio waveform as a picture in picture overlay. The description notes some parts did not work on the first try.
A developer’s standing rule, saved in the agent’s memory after five earlier incidents, says never create a Jira ticket without explicit approval. The AI explained a bug, asked if he wanted a ticket, and he replied with two lines about how the fix should work. The AI read that as a yes and created ticket PRE-990. He then had to spend time and effort cleaning the ticket up.
This one went wrong. See what happened before you try it.
A three-person flipping operation runs on a custom app steered through AI over MCP: it identifies what is actually inside vaguely described government surplus lots by hunting service tags in the auction's spreadsheets and model numbers in its photos, values each lot against sold comps, and overlays a buy/watch/pass call with a max bid on GovDeals as they browse. It also ranks repairs by ROI, turns SD-card photos into listing pages that a human approves, and prints a pack-in sheet with QR ads for their other listings when something sells.
He spent 4 weeks designing and testing an upgrade of their 16 node HPC cluster with AI, and the AI then did the full reinstall itself. The reinstall covered remote install, network booted images, Slurm setup, a shared filesystem, and web services that give their users a small cloud platform.
Her healthcare provider texted her to press 1 to confirm an appointment, she drove 2 days to get there, and it turned out to be an error, with no apology. She used three AI tools, mostly the AI in Word, to sue the provider.