Update
Decisions the agent now asks during a run: a budget the platform refuses, an account it cannot pin down, which ad to recreate
TL;DR
Decisions about money, accounts and creative references are now asked mid-run instead of guessed. Before, a budget below a platform's minimum ended the task as a failed step — and an automatic argument repair could raise the value itself.
The three decisions
- Money the platform refuses — a budget below its minimum, a bid it requires, a cost cap. The options offered are the minimum the platform named, automatic bidding where that applies, or skipping the step.
- An account the run cannot pin down — the candidates it found, or every connected account of that platform.
- Which spied ad to recreate — the top candidates, numbered as in the results table, plus "search again" and "no reference". Not asked when your goal already names the ad or tells the agent to choose.
How it behaves
- Asked even with auto-approve on, whenever a person can answer: a run a person started, or a scheduled run that has an approver and auto-approve off. Never on a run started by an API key or by email. Where nobody can answer, the value is not guessed — the step is left undone and the report says why.
- Your answer becomes a re-plan, not a resumed step with the old arguments. "Skip" resumes the rest of the run.
- At most three questions per run, never twice for the same step.
- The automatic argument repair no longer touches money or an ambiguous account. A model fixing a refused call is fine for a malformed field; it is not fine for raising a budget.
Frequently asked questions
- Will the agent raise my budget to meet a platform minimum?
- No. When a platform refuses a budget as too low, the run pauses and asks you, offering the platform's stated minimum, automatic bidding where it applies, or skipping the step.
- Does the agent ask questions when auto-approve is on?
- For three decisions, yes — money a platform refuses, an account it cannot identify, and which spied ad to recreate — whenever a person can answer: a run a person started, or a scheduled run with an approver. Runs started by an API key or by email are never paused for a question.