Yesterday I was on the phone with an old friend, and somewhere between catching up and complaining about work, we started comparing AI agents to prototypical coworkers. One was the eager intern who ships fast and breaks things. One was the cautious type who reads every instruction literally. Then we got to Astra, OpenAI's newest model, and we both laughed, because we'd had the exact same experience: Astra is the brilliant guy from the third floor who knows everything and wants to do none of it.
Hand it a task and it answers with a set of thoughtful questions. Have you considered the timezone edge case? What about legacy clients on the old API? Should we check with legal first? Each question makes sense. Each question is smart. But seen together, they are a refusal to work, dressed up as diligence.
The joke landed because it's true. Models have stopped being interchangeable. They have temperaments now. For the last two years, everyone has been benchmarking them like CPUs, comparing obscure benchmark scores. Wrong frame. A benchmark tells you what a model can do. It tells you nothing about what it will do on a Tuesday afternoon with a half-specified ticket. That gap is personality, and personality is a staffing question, not a procurement question.
So treat it like one. You wouldn't put your most anxious senior engineer on a fire. You wouldn't let your fastest junior near the billing system. Match the model to the work, the way you match people to work. And if you're in a larger organization, you have to make sure people start asking "Which model would you hire for this task?" more often.
By the end of the call, we had agreed on the only management technique that works on Astra. It's the same one every team lead eventually learns: kick it in the shin and say "start."