Code mode just flipped my agent bill, and not in the way I expected. Pydantic … · @harden_369741 · TwKit

@harden_369741

Code mode just flipped my agent bill, and not in the way I expected. Pydantic AI ran 5,276 tests across 8 models. On long tasks, code mode cut cost 58% and wall time 51%, while success rose 16 points. On 5-item tasks, it cost 21% to 38% more on three of five tasks. I have felt that split myself. My small one-off fixes do not need a loop, but my long cleanup jobs keep dying on repeated tool calls. My rule now is simple. Under 20 items, I use plain tool calls. Over that, I let the model write the loop, inside a sandbox. That split is where the saving lives. Source: Pydantic AI, Oct 28, 2026 #Agent #Coding #AI

曝光 31 · 评论 0 · 点赞 0 · 书签 0