@Leechael: The new generation of models should no longer use subagents. Subagents aren't as effective as having sessions exchange …

X AI KOLs Following 新闻
subagents ai-models synchronization token-efficiency gpt-6 codex

摘要

The article discusses the inefficiency of using subagents in AI models, suggesting that direct message exchange between sessions might be more effective, based on observations about GPT-6 Astra's token usage in Codex.

The new generation of models should no longer use subagents. Subagents aren't as effective as having sessions exchange messages with each other for synchronization.
查看原文
查看缓存全文

缓存时间: 2026/09/10 12:30

The new generation of models should no longer use subagents. Subagents aren’t as effective as having sessions exchange messages with each other for synchronization.

Gangqiang Yao (@yaogangqiang): 一直感觉 GPT-6 Astra 用 subagent 额度掉得特别快，就去翻了下 Codex 源码。

发现一个有问题的地方：主 agent 等 subagent 的时候，默认每 30 秒醒一次问「好了没」，没好就再睡。每问一次都是一次模型请求，有些浪费。

有点像 leader

相似文章