ETF Agent Details Runtime Reminder Coordination
The ETF Agent team described an architecture that turns changing runtime facts into timely advice for models while leaving facts, execution authority, and delivery to their existing modules. The design separates persistent status from repeated action suggestions, deduplicates reminders by type, and records delivery only when a reminder actually enters the submitted model request. The team says these paths have been validated in a controlled setting, but its effect on real research performance remains unobserved.
Tracking whether reminders actually reach the model provides a more reliable basis for recovery, deduplication, and runtime guidance than treating reminder generation as delivery.
一项研究任务进行到一半，Agent 正在比较几只基金的持仓和费用。它已经读过一些资料，尝试过计算，也开始整理答案。
这时，运行环境掌握着一些模型未必能自己准确判断的事实：可用请求正在减少，上下文越来越拥挤，一段已生成的正文接近修改次数上限。系统还发现，研究已经持续了较长时间，或许值得停下来检查证据缺口。
这些信息都值得告诉模型，但问题随之而来：哪些提醒应该出现在这一轮？它们有没有真的进入模型请求？提醒模型“考虑复核”，是否意味着系统已经批准再调用一个 Reviewer？
我们在 ETF Agent 中设计行动提醒协调时，核心判断是：提醒负责把运行事实翻译成适时的建议；事实、执行权限和最终交付，仍由各自的原有机制负责。
这听起来像是在 Prompt 里加几句话。真正需要处理的，却是提醒的来源、组合、容量、送达记录，以及它不能越过的职责边界。