Hark vs Instinct: Dinner analysis @adcock_brett s Hark took 20 minutes to order me a pizza and got it wrong. Two pizzas, wrong restaurant. @noahrshinn s Instinct emailed 50 restaurants to organize a private dinner for 20 people and got it right. That’s backwards. Finding a restaurant requires research and reasoning. But once you’ve identified the restaurant and what you want, ordering should be deterministic. Select the items, confirm the details, complete checkout. This is a 5 min process at most. Finding a private dinner for 20 is the opposite. Open-ended research, dozens of venues, inconsistent replies, contracts, and fine print. Instinct researched 50 venues, emailed each one, followed up on unanswered requests, and organized 41 responses into one comparison sheet. It caught a contract quoting 20 guests but charging food for 15. It flagged “private” rooms that were actually curtains or mezzanines. It asked for approval before sending each batch. The simple task failed. The complex one worked. Agents are becoming remarkably good at open-ended research and communication, but still struggle with routine execution. Once the reasoning is done, execution shouldn’t require more reasoning. It should follow defined, permissioned workflows. That’s an infrastructure problem, not just a model intelligence problem.