News for hackers. Stuff that matters.
Re:In 27,280 coding agent runs, 64% of inference spend went ... by ricardobeat on Wednesday November 18, 2026 @05:48AM (#50544618)
Surprised that Crush [1] was not included in the benchmarks. It’s my default harness for open models, and with the latest subagents update very efficient for orchestration. The built in LSP support helps models a lot in reducing the amount of tokens spent on code syntax errors.

[1] https://github.com/charmbracelet/crush

Re:In 27,280 coding agent runs, 64% of inference spend went ... by andreynering on Wednesday November 18, 2026 @07:53AM (#50545734)
Thanks! I help maintain Crush, and sub-agents are indeed pretty fun to use, and also very useful of course.