this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks.
「简单题不影响，难题可能变差」——这个不对称性极为危险。它意味着我们在用简单任务验证 Agent 可靠性时，得到的是虚假的信心。而当 Agent 真正面临高风险、高复杂度的任务时，上下文累积已经悄悄关闭了它的自我验证模式，在没有任何预警的情况下退化为浅层推理。这是一种「隐性能力衰减」，比显而易见的失败更危险。
Vicky Zhao indirectly frames the answer for "why have a zettelkasten?", especially for learning, as overcoming the "illusion of competence" which is closely related to the mere-exposure effect and the Dunning–Kruger effect.
Califf, Robert M., Adrian F. Hernandez, and Martin Landray. ‘Weighing the Benefits and Risks of Proliferating Observational Treatment Assessments: Observational Cacophony, Randomized Harmony’. JAMA 324, no. 7 (18 August 2020): 625–26. https://doi.org/10.1001/jama.2020.13319.
Because they're more integrated and try to serialize an incomplete system (e.g. one with some kind of side effects: from browser/library/runtime versions to environment to database/API changes), they will tend to have high false-negatives (failing test for which the production code is actually fine and the test just needs to be changed). False negatives quickly erode the team's trust in a test to actually find bugs and instead come to be seen as a chore on a checklist they need to satisfy before they can move on to the next thing.
So finally I'm coming out with it and explaining why I never use shallow rendering and why I think nobody else should either. Here's my main assertion:With shallow rendering, I can refactor my component's implementation and my tests break. With shallow rendering, I can break my application and my tests say everything's still working.This is highly concerning to me because not only does it make testing frustrating, but it also lulls you into a false sense of security. The reason I write tests is to be confident that my application works and there are far better ways to do that than shallow rendering.