1. Provisioning, once. On its first run the agent looks in your Google Drive for a workbook called "CogniAgent Recruiting Data". If it is not there it creates it, adds the six tabs it needs with headers and a few sample rows, and removes the empty default sheet Google adds. Run it again (or install it again) and it reuses the same workbook rather than making a second one. 2. One candidate per check. Every five minutes candidate screening AI agent takes the oldest application marked for screening and claims the row before doing anything else, so two runs can never process the same person twice. 3. The questions. It emails the candidate the screening questions for the job they applied to, taken from the Jobs tab. It asks only what the job requires, and it will not ask about protected or sensitive characteristics — if a rule you wrote names one, it refuses that rule and tells your escalation address why. 4. The decision. Answers are scored against the ScreeningRules tab: required skills and work authorisation are hard requirements; start date and pay expectations are noted for a human rather than used to reject anyone. The result is written back with the reason in plain language. 5. A human sees anything unclear. Borderline results go to a recruiter for review before the candidate hears anything. Out of the box, every rejection is confirmed by a person — automatic declines are off until you turn them on. 6. Silence is handled. If a candidate never replies, the agent follows up, then records "no_response" and tells the recruiter. It does not leave the row waiting forever.