Keeping raw logs and search results inside the program, with only printed outp… · @yinpengxuan · TwKit

@yinpengxuan

Keeping raw logs and search results inside the program, with only printed output going back to the model, is the part I'd copy into any harness. So much agent context gets burned on a full test log when three lines of it mattered. The detail I like is that full results still get logged for later search, so trimming what the model sees doesn't cost you the debugging trail.

曝光 1 · 评论 0 · 点赞 0 · 书签 0