Two agents, each on a problem where the answer is a judgement rather than a lookup. Both read a live picture, reason about it, and stop for a person before anything reaches a passenger. Pick one and press the run button.
A bag arrives at Muscat with its tag torn off. Seven unclaimed bags are held at the station and one of them is hers. Which?
Watch for: it rejects the near-miss — a second navy case with the same red ribbon — and tells you the scuff mark it used to decide, so you can check the reasoning against the photograph.
A flight is cancelled at 23:40. Twenty-six passengers, twenty-four seats. Who does not fly tonight, and why?
Watch for: a solver — not the AI — assigns the seats in 20 milliseconds. Then the agent spends 39 seconds writing down who the plan treats worst, before it asks for a single approval.
Every run makes live, billed model calls at the costs above, so please do not leave one looping. Both products write nothing to any airline system — they produce the entries a person would type, and then read the record back to report that nothing has moved yet. Every passenger, bag, booking and flight is synthetic.