LobeHub agents review and extend their own coordination code
LobeHub's lobehub-coder agent and OpenAI's Codex exchanged code-review feedback on merge request #21852 for 18 minutes without human involvement, before a human engineer merged the change into the main branch. The proposed feature lets one agent specify which model another agent should use when assigning work. 新智元 also tested an overnight Claude and ChatGPT editorial workflow on LobeHub, finding that agents retained user feedback for later runs, while configuration, authorization, and some handoff steps still required manual work.
The reported workflow provides a concrete example of AI agents reviewing code that improves the infrastructure they use themselves, with humans retaining final approval and acceptance responsibilities.
10月31日凌晨4点04分,开源项目LobeHub的GitHub页面上,一个名为lobehub-coder的账号,跑到Codex审查机器人提交的代码审查意见下面留言,意思是问题已经修复。
在随后的18分钟里,双方来回交流了好几轮,lobehub-coder先修复了Codex两天前指出的两个问题。紧接着Codex又提出三处新问题,它修复了其中两处,到最后一处却没有照办,而是回复说这个问题确实存在,但经过一次独立评估,目前的处理方式可以接受,随后逐条说明了理由。
整个过程没有任何人类参与。彼时,员工们绝大多数还在睡梦中。
两个AI讨论的这段代码,来自lobehub-coder在10月29日凌晨提交的一个合并请求,编号#21852。它要实现的功能是,让Agent在给其他 Agent分配任务时,可以顺便指定对方使用哪个模型,就像工头安排工作时多交代一句,这件事不着急,交给便宜一些的机器去处理。
#LobeHub#lobehub-coder#OpenAI Codex#Acceptance#DeepSeek Harness#Claude Code#ChatGPT