ETF Agent separates agent workloads across three processes
An ETF Agent engineering article describes a three-process architecture comprising Core API, Product API, and Background Worker, giving product management, foreground research, and background execution separate survival commitments. Controlled tests found that API restarts could leave an accepted background task running, queued foreground requests could take precedence over unstarted background work, and an unavailable Product service did not prevent account-scoped research and result persistence. Validation was limited to same-machine processes, isolated PostgreSQL, and a synthetic model, with no production reliability, latency, or cost measurements.
The architecture offers practical guidance for separating interactive and long-running agent workloads while retaining existing task, run, provider, and publication authority, but its production benefits remain unquantified.
本文结论
- 三进程首先让产品管理、普通研究和后台执行拥有不同的存活承诺;受控实验证明 API 重启不必重启原后台 Task。
- 拆宿主不能拆业务权威;Task、Run、Provider 和 Publication 的原始身份、资源与结果继续由各自协议负责。
- 新后台任务对已登记的同组前台请求让位,但已经启动的工作不抢占;原位容量交接避免释放再领的空隙。
- 租约过期不等于旧 Worker 或物理请求已停止;本地端口仍可监听也不等于原 Product 存活。
- 验证限于同机进程、隔离 PostgreSQL 与合成模型;尚无生产可用性、延迟或成本改善的量化结论。
适用
- 同一 Agent 同时提供交互问答、订阅管理与定时后台工作,且需要分别维护这些工作负载。
不适用
- 没有独立后台存活需求的单次问答应用,或需要跨机器自动恢复、不同版本滚动发布的系统。
#ETF Agent#Core API#Product API#Background Worker#PostgreSQL