AI Agent
It must work, and it must stop.
In one sentence
An agent needs a goal, tools, a place to stop for a person, and a record when it’s done.
The common enterprise mistake is treating a chat box as a hired employee. A chat box produces text. An agent also changes state: publishing, tickets, link checks, ledgers. Changing state means errors. Write the boundary first.
Good work for an agent is repetitive, inspectable, and cheap to get wrong. Bad work is production with no permission table, no replay, and no human signature. That is how I decline projects.
Xiaoyuan’s browser helper is a narrow example: it acts in a locally logged-in account and does not upload cookies. Narrow can still be useful.
The longer argument is in the essay on the real boundary of agents in companies.