Partner's user

"What would you do today?" runs the marked readings against today's data and shows the rows the agent would see, with the source and the link to the proof. It does not save, does not send, does not execute any action and does not appear on the agent's scoreboard. Two important readings of the result: if a source comes back "no data today", the agent's report would come out blank in that part; and if the rehearsal reads nothing, it does not prove that the agent works — the screen says so instead of giving an empty ok.

The screen became the standard list: search, filters, saved views, columns, export and trash with Undo. Click the record (or New) to open the form, with the detail tabs when there are any.

The path has four stops and none is optional: interpret (nothing is saved), review field by field, rehearse if you want, and only then create. The "Create the agent" button is the only one that puts something live. A single button that created directly from the text would be the same path with the review implied — and an implied review is a review that did not happen.

If your request mentions something that does not exist on the platform, it goes to the "this I do not know how to do" list, with the name you used and the reason. It is not silently ignored (you would think you had asked, and the agent would never do it) and it is not swapped for the capability with the most similar name (the agent would do something else, with the authority of having been asked). This list is also useful in another way: it is the record of what customers ask for and the platform does not have yet.

"Every 5th", "every Monday", "when it is more than 10 days overdue" are read by deterministic rules. What no rule can read is declared in "this part of your request I did not understand" and the agent is born running only on request. It is on purpose: a schedule wrong by one day is an agent that runs on the wrong day forever, without ever raising an error — the routine runs, the report goes out, the customer receives it, just on the wrong date. Things like "first business day" and "end of the month" also appear there, because they depend on the holiday calendar and the length of the month.

Writing a request in plain language is not authorization to act. The agent created here writes the suggestion and touches nothing; the actions you authorized come in turned off. Raising the level (to REVIEW or EXECUTE) and turning on the power are separate acts, in Agent Autonomy and in the Transparency Center, where who did it and when is recorded. Editing the request later creates a new version: the previous one is kept, because it is what explains the decisions the agent has already made.

In the version history, "Revert to this version" republishes the old content as a NEW version — nothing in the history is deleted, and decisions already made remain explainable by the version that made them. Powers added after the target version are deactivated (not deleted): turning them back on is your decision, in the Transparency Center.

Write it the way you would explain it to a colleague who started today: what to check, when, and what you want to receive. Examples that work well: "every 5th, check the contracts without adjustment and send me the list"; "when a receivable is more than 10 days overdue, notify me with the customer's history"; "every Monday, the week's cash summary". The more you use the name the platform uses for the subject (receivables, contracts, orders, tickets, budget), the better the proposal.