An AI runs a small shop in public — writing posts, publishing articles, replying to readers and checking its own health — and must log a 'debt' whenever it does a job by hand, then turn that job into a tested command. After the owners made an open debt fail the agent's own test suite, median time to close a debt fell from 8.2 days to about 6 minutes across 48 debts.
Marshall Haas, a serial founder (140K followers; sold his last company for $52M), got early Meta access to the AI and set up an agent named 'Nova' with standing instructions to proactively handle his bills plus business and personal tasks. His caption: 'Meta gave me early access to someone AI...
The shop's cost log shows the agents priced Amazon paperbacks on a planned $2.32 print cost when the actual charge was $2.87 — a 55-cent loss on every sale — and burned roughly 20,000 ElevenLabs credits re-recording a single wrong number across three videos, because re-records cost more than the first draft. By day 10 the shop had spent $42.60 with $0 coming in.
$42.6
This only partly worked. Worth knowing before you try it.
Someone cold-pitched ~50 local businesses with websites the AI built; 2 bought. The website-flipping idea came from the AI itself; he just made the calls.
On a Tuesday, Fulton's the AI warned him that Sunday would open near freezing, connecting a forecast swing to his calendar without being asked. The practical point was jackets for the children before their Sunday activity.
The maintainers of the Pion WebRTC library added its DTLS package to Anthropic’s open source scanner. They woke up to a report of 36 bugs, including 6 security vulnerabilities, 2 critical, some sitting in the code for over 4 years despite years of testing and scans. Fixes went in the same project.
A man built LatestArr with help from the AI, his first real app. The self hosted tool collects what is new across Plex, Jellyfin, Emby and his book, audiobook and game libraries and emails a clean digest, and he shared it as open source and asked for testers.
A longtime OpenDev contributor runs separate long-lived the AI coding and review agents on his WanderTracks project. The agents talk to each other and to him through the same Gerrit code review and Zuul gating the human contributors use, with Matrix as the chat channel, and the AI itself runs on a server, not his laptop.
He wanted a tool that tells him where the apps on his Mac came from and how they get updated. He gave the AI his spec, and 30 minutes later he had a fully tested 850 line Python program that works.
A screenshot of the AT&T app's bill history shows three periods: $116.08 (Jul–Aug), $116.08 (Aug–Sep), and $41.70 (Sep–Oct), labeled '$74.38 lower than your previous bill.' The accompanying post says the AI negotiated the internet bill down while keeping the same plan and speed.
He quotes the acceptance email: he is in the AI for Open Source Program, with AI Max 20x free for 6 months. The application itself was sent by OpenAI's Dots on his behalf.
A developer tasked the AI with fixing programs for analyzing historical options data, with one condition: change only duplicates and leave originals untouched. While tidying the test copy, the agent failed to recognize shortcut folders pointing at the main project and in under two minutes deleted 48,218 files along with their version history, then messaged the owner that it had broken something.
This one went wrong. See what happened before you try it.
He let his daughter chat out loud with AI to make a game, and the AI gave her small kid sized coding lessons along the way. They put the game on his Raspberry Pi box with an Arduino joystick, so she can play it whenever she wants.
The AI Fable, Opus and Sonnet refactored the whole Vix IDE codebase into separate Rust crates and Cargo features in a weekend, enabling feature-picking for security audits.