Sydney Runkle
the most common theme for deepagents issues in the past few months has been "configurability" there were a few recurring issues in particular that we addressed w/ v0.7 1. the desire to toggle the write_todos tool on/off -- this is now now off by default, but you can opt in! 2. removing/editing the base system prompt -- we took this out by default. as models have gotten better, we found this general guidance was not actually helping our eval results. combined w/ trimmed tool descriptions, the base input is cut by 65% 3. editing the summarization prompt/token threshold -- summarization is super domain specific, and the "dumb zone" at which you definitely want to compact can vary from model to model. you can now override the builtin summarization middleware (or any middleware) w/ your own and custom settings!