Here's another cool piece about smart knot management using a harness, which allows the model to not work with a long knot. In recent context language models, the model itself rewrote its context as a file on disk. The harness is responsible for that here. The result is cool: the quality usually does not sink (and sometimes even rises). There really is a problem with resetting caches and most likely the lack of financial savings as a result. The authors of the article did not publish code artifacts, I asked Claude to compile a skill that reproduces the approach described in the article, and I published it here: it teaches the agent two things from the article: not to load large files in context, but to search with grep and read in limited windows, and during a long mission, to maintain the same six-part delivery summary (goal, verified facts separately from hypotheses, changes made, running tests, deadlock, planning steps more). The skill is collected by itself in long multi-step tasks and large inputs, and context reset is left to cloud code - automatic compression according to its threshold, /compact or /clear hands; In any case, the plugin hook returns the brief, and the agent continues from there, rather than digging through everything again. The result is EoCC; SAC (reset after each operation) cannot be done in cloud code, it is only in the reference harness in turnip - along with EoCC and uncompressed baseline. I haven't tested it thoroughly yet, there may be bugs. PR accepted :) You don't need a long context Peter Blackack, Saurabh Muralidharan, Jindong Jiang, Yonggen Fu, Wenmin Baiyun, Hanrong Ye, Ryan Angeli, Ali Tagivabakhshi, Shi-Yang Liu, Siro Ouyang, Greg Heinrich, Pablo Molchanov Paper: Review: Code: N/A (unofficial reimplementation) Developer: Language Models Handle Long Tasks based on agents and large documents in a short context window without loss of accuracy.