ETF Agent reconnects delayed human replies to the original research run
ETF Agent authors describe a persistent Human-in-the-Loop design that separates request acceptance, response receipt, response adoption and effect commitment, with the original run recovering after a delayed reply. Their open ETF Agent v0.13.0 example persists the question, records the reply as received but not yet adopted, and requires an explicit resume so the original call consumes it. The article uses synthetic inputs and says its local implementation does not establish production reliability or that real models will ask appropriate questions.
The report offers a concrete pattern for long-running agents that must preserve request identity and reconnect a late human response to the exact run and consumer that requested it, rather than treating receipt as completion.
研究 Agent 正在撰写一份基金风险报告，读过持仓摘要，也找到了基金披露资料，却缺一个决定分析口径的问题：“这次比较以哪个基准为准？”它可以先整理已经取得的资料，同时把问题交给用户。用户可能正在开会，过几个小时才回答；期间页面已经关闭，服务也可能重启。
难点不在于画一个“请确认”的弹窗，而在于：回答被系统收到之后，究竟有没有进入最初需要它的那次研究？ 更棘手的是，“我同意”可能只是回答问题，也可能是对受保护操作的审批。把它们混成一个布尔值，Agent 就会失去解释自身行为的能力。
本文以一个长期运行的研究 Agent 为例，解释持久 Human-in-the-Loop（HIL）如何实现非阻塞提问、精确回复、同 Run 恢复与有界审批。例子均为说明机制的合成输入，不涉及真实账户或持仓。
一、先分清四个时点
传统聊天框很容易把“问题发出了”“用户看到了”“用户回答了”“研究已经采用回答”合并为一个已完成状态。但它们是四个可独立失败的事实。
#ETF Agent#ETF Agent HIO Stage A–D#Human Interaction Owner#Session v3#open ETF Agent v0.13.0