EXPERIMENT XI · AI INFRA

离执行更近Closer to Execution

让前端直连沙箱执行，多轮任务能省下多少等待？When the frontend connects directly to the sandbox, how much waiting does a multi-turn task save?

待开始Not startedE2BE2BWebSocketWebSocketPrompt CachingPrompt Caching
  1. Q提出可证伪的问题State a falsifiable question当前阶段Current stage
  2. R做最小运行路径Build a minimal runnable path未开始Not started
  3. E连接真实验证Validate against reality未开始Not started
  4. P公开证据与限制Publish evidence and limits未开始Not started

为什么值得做Why it is worth doing

把 Agent 执行从集中调度的 Worker 挪进沙箱，拆开控制面与执行面，文件操作和命令就在项目环境里完成。省下的时间需要被量化。Moving agent execution from a centrally scheduled worker into a sandbox, and splitting control from execution, puts file operations and commands inside the project environment. The time saved needs to be quantified.

Q

假设Hypothesis

HYPOTHESIS

在多轮编码任务中，WebSocket 直连沙箱比经由 Worker 转发减少明显的往返与准备时间，且收益随轮次增加。In multi-turn coding tasks, a direct WebSocket sandbox connection cuts round trips and setup time visibly compared with routing through a worker, and the gain grows with each turn.

R

最小装置Minimal setup

MINIMAL SETUP
  1. 01用开源组件搭一个最小对照：Worker 转发 / WebSocket 直连 E2B 沙箱Build a minimal comparison with open-source components: worker relay / direct WebSocket to an E2B sandbox
  2. 02准备 5、10、20 轮的编码任务Prepare 5-, 10-, and 20-turn coding tasks
  3. 03记录每轮往返时间、准备时间和失败重连Record per-turn round-trip time, setup time, and failed reconnects
E

什么会证明我错了What would prove me wrong

FALSIFIED IF

直连方案在多轮任务中没有稳定的时间收益，或失败率更高。If the direct approach shows no consistent time gain on multi-turn tasks, or fails more often, it is not justified.

P

将会公开What will be published

WILL PUBLISH

对照实现、测量脚本和逐轮时间线。The comparison implementation, measurement scripts, and per-turn timelines.

前期线索Early leads相关的已有工作，不是这个实验的结论Related earlier work, not a result of this experiment

  • 执行架构迁移经验Execution architecture migration在 Eazo 参与过执行架构向沙箱的迁移；这里用开源组件独立复现。Worked on migrating the execution architecture into sandboxes at Eazo; this reproduces it independently with open-source components.

实验日志Run logRUN LOG

还没有记录。实验开始后，每一次运行、失败和修正都会记在实验室里。Nothing logged yet. Once the experiment starts, every run, failure and fix will be recorded in the lab.

去实验室Go to the lab ↗