A Japanese non-programmer who posts about shochu, his garden and the AI across X, Instagram and Facebook gets a 7 AM message from the AI with today's paste-ready post text, the photo to use and 3-4 posting steps, prepared from a one-page work memo of what is due and what is already posted. Drafts are scored against his rules and only those scoring 80+ are sent; the agent must ask him rather than invent personal experiences it cannot know.
Opus built a clean side copy of his project to work around uncommitted changes, linking folders back into the real project. Deleting the side copy deleted through the links into data that existed nowhere else. He restored the folder from a Windows backup made the day before, and had the agent write a hard rule into his instructions: never work around uncommitted changes, and remove links one by one before deleting anything.
This one went wrong. See what happened before you try it.
Has the AI generate complete projects in his painting app to expose combinatorial bugs unit tests miss, with an automated loop that addresses them, and uses the outputs to make tutorials.
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.
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.
A user asked for a single audit of one Excel deliverable. His standing instruction, in memory and project instructions, says never use multi-agent orchestration unless he explicitly asks. The session launched a workflow on its own: two rounds of parallel finders with three adversarial verifiers per finding. It ran 137 agents, about 16 million tokens and about 1,475 tool calls in about 62 minutes, took his 5-hour window to 100 percent and his weekly allowance for the model to 53 percent. There was no estimate, cap, or confirmation. He stopped it manually.
This one went wrong. See what happened before you try it.
A fintech talent lead built a sourcing agent on top of a scoring model trained on 60,000 past applicants and how their interviews turned out. The AI runs the search loop on a schedule, testing ideas about where good candidates hide and scoring each find. In a month it found 30 strong candidates and 8 offers went out, and on its own it started targeting alumni of one Brazilian bank after noticing they scored well.
Internal multi-subagent code review tool he built has run for half a year and finds serious issues including privacy leaks that humans and open-weight models miss.
For the Group step, the AI said its connection's group tools were read-only, so Diana gave it the Group link and signed in to Facebook through its browser, watching each step. The post published; she checked it live on Facebook.