Why this matters more in the agent era

In a single chatbot turn, you can often get away with a sloppy context, because there is only one exchange. Agents are different. An agent runs in a loop, and a typical task in a system like Manus uses around 50 tool calls. Every one of those calls produces output that, by default, piles up in the context window. Without active management, the window fills with old tool results, and the model's quality starts to degrade. That degradation has a name, context rot, and it is the subject of the next chapter.

This is why the articles treat context engineering as the number one job of an engineer building agents. The reasoning quality of the underlying model is becoming a given. What you control, and what decides whether your agent feels magical or broken, is the context you build around it.