@sweis The 1,024-token context cap is the part I'd most want to see move next. Most agent requests blow past that before the model even starts answering.
@addyosmani Writing down what you expect the diff to look like before the agent runs is the habit I'd keep. Skip that and it's way too easy to grade the agent's summary instead of the code.
@OpenAI Curious what share of the findings turn out to be real, exploitable bugs versus noise. That ratio decides whether smaller teams can afford to run something like this.
@bindureddy The 20x only holds if someone actually owns the retraining every time the base model moves. That upkeep is the line most teams forget to budget for.
@rauchg Agree. A fresh box every run means reinstalling deps and re-reading the repo before any real work starts. Carrying auth along safely seems like the harder part than memory though.
@thsottiaux@tnm Curious what broke first once agents ramped up. Was it the sheer number of pushes, or fetch load from every agent cloning the monorepo?