你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。
最容易做错的，是bojieli / ai-agent-book；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：Agent成败，模型只占三分之一。
The sentence worth stopping for is this: in the book's framing, the model is only one third of whether an agent works.
The README states the formula directly: agent = LLM + context + tools. LLM here just means the language model itself. The point is not that models do not matter. The point is that model quality gets 1 of 3 votes, not 3 of 3.
Chapter 2 makes the sharper claim: context sets the ceiling. In plain English, the instructions, memory, and retrieved documents around a model can let a decent model beat a top model that is working blind.
Chapter 1 adds the part many teams skip: the stronger the model gets, the more the surrounding harness matters. Here, harness means the checks, validation, and correction loop around the model. Better raw ability does not remove the need for guardrails.
The line worth keeping is simple: don't judge an update by how many features it lists. Judge it by whether it changes your next decision.
My boundary is narrow. This takeaway comes from the current README plus Chapters 1 and 2, not from a fixed benchmark on one stack. So the safe conclusion is not "models are overrated." It is "model choice is only part of the system."
If you know someone about to spend weeks swapping models before fixing context or validation, share bojieli / ai-agent-book with them.
真正该讨论的是：bojieli / ai-agent-book