📄 中文摘要
在构建自主 AI 代理时，开发者常常会遇到 API 数据膨胀的问题。尽管初期表现良好，但随着使用时间的增加，AI 代理可能会忘记核心指令，开始出现幻觉，导致令牌使用量激增。这是因为 API 设计主要是为了传统软件，而非大型语言模型（LLM）的上下文窗口。文章指出，解决这一问题的关键在于优化 API 调用，减少不必要的数据传输，从而提高 AI 代理的效率和稳定性。
API 数据膨胀如何毁掉你的 AI 代理（以及我如何在 Python 中将令牌使用减少 98%）
📄 English Summary
How API Data Bloat is Ruining Your AI Agents (And How I Cut Token Usage by 98% in Python)
When building autonomous AI agents, developers often encounter the issue of API data bloat. Initially, the agents perform well, but over time, they may forget their core instructions and start hallucinating, leading to a spike in token usage. This problem arises because APIs are primarily designed for traditional software, not for the context windows of large language models (LLMs). The key to addressing this issue lies in optimizing API calls and reducing unnecessary data transfer, thereby enhancing the efficiency and stability of AI agents.