为什么别急着加第五个 AI 助理:多智能体交接会丢上下文
作者主张与其堆多个助理互相交接,不如用单个 Grok Bot 在沙盒里保留完整上下文跑 100 步,成功路径存成 skill 复用。
WHY ADD A FIFTH AI AGENT IF EVERY HANDOFF LOSES PART OF THE TASK? Every handoff turns the original files into a shorter summary. By the time the last agent hits an error, the detail it needs may already be gone. I’d start with one Grok Bot in a sandbox instead. Give it the task context as files and 100 steps to work through them. It can read the source, try an approach, inspect what failed and try again without asking another agent to reconstruct the story. The useful part comes after it finds a route that works: save those steps as a skill, so the next similar question doesn’t start from zero. That’s the loop in the video: read → try → hit an error → check the evidence → retry → save what worked Before adding another agent, I want to see what one agent can finish when it can keep the full context.
这条是引用着另一条写的:x.com/shiqway92/status/2109953296232267810