ETF Agent vs. General Agent Frameworks
An engineering analysis compares ETF Agent with Pi, Claude Agent SDK, and LangGraph across execution, permissions, recovery, artifact delivery, and financial research responsibilities. It does not publish a scorecard because the four systems were not tested with the same models, data windows, budgets, and question sets. Drawing on the frameworks’ official mechanisms and checked ETF Agent implementation evidence, the analysis identifies financial methods, provenance, permissions, result identity, and verifiable delivery as the system’s distinctive strengths. It argues that its custom runtime should continue to be tested against the mature alternatives rather than treated as necessary simply because it exists.
The analysis separates reusable execution mechanisms from financial-domain responsibilities and shows why permissions, source identity, recovery, and artifact delivery must be evaluated directly. It also demonstrates how to compare agent architectures without presenting unverified benchmark rankings as established results.
` 同步校注（2026-11-04）：文中涉及撤权、来源消费和恢复的历史案例，应与当前 Run 授权生命周期区分。当前实现中，Host 已授予的权限在同一 Run 及其恢复中持续有效，不因外部权限表变化重新撤销；新 Run 独立准入。账户、scope、来源范围、内容身份、预算和显式取消仍按原合同执行。历史撤权复现用于解释当时的问题，不代表当前 Run 内重新授权机制。
工程篇｜沿一条基金研究任务，检验自研系统到底增加了什么，又承担了什么。
“为什么要做金融垂类 Agent”可以在原则层回答，但工程团队最终还要面对更具体的问题：既然 Pi、Claude Agent SDK、LangGraph 已经提供了大量执行能力，ETF Agent 为什么还存在自己的运行时、结果存储、发布和恢复机制？哪些是合理差异，哪些只是需要持续偿还的工程成本？
本文不做四个产品的分数排行榜。我们没有让它们在相同模型、资料窗口、预算和题集下完成一次横向实测，因此不能给出胜率、金融准确率或成本排名。比较采用两个坐标：通用框架官方提供的机制，以及 ETF Agent 已核对实现与原始验收能够支持的具体行为。