Architecture for Persistent, Long-Running AI Research Agents
The article argues that long-running research requires more than larger contexts or longer timeouts: systems must separately manage the user task, each execution Run, working context, and saved research artifacts. It recommends preserving evolving judgments and execution receipts, delivering staged results, recovering interrupted Runs without duplicating effects, and treating explicit continuation as a new Run linked to its parent. A synthetic fund-comparison scenario illustrates these design patterns but does not constitute a completed investment study or verified trading capability.
The framework gives AI agent builders concrete guidance for preserving evidence, user instructions, execution identity, and deliverables across context limits, interruptions, and later continuations.
用户交来一个问题：“请比较两只主题基金与我的组合，先给我一份有依据的初稿，再把关键疑点查清。”研究可能需要多份文件、几轮计算和不同方法的复核。中途用户会离开页面、追加条件，服务也可能重启；资料越积越多，单次模型请求装不下全部历史。
这里的“超长”没有统一的小时数或 token 阈值。它指的是：一项研究的生命期，已经超过一次请求、一个上下文窗口或一段不间断进程能够承载的范围。
把上下文开大、把超时调长，可以延后某些限制，却无法单独解决任务归属、原始依据、重复执行、结果保存和下一次怎样继续的问题。
真正支持超长研究的系统，需要让研究判断持续演进，让执行事实稳定保留，让用户在每个有价值的阶段取得成果。 本文沿开头的教学任务展开这些关系；示例不是一次已完成的真实投资研究，也不代表自动交易能力。