... honestly, the speed is crazy. But the annoying part is when the agent understands the task, but not the history of the product. Example: it can build a new feature nicely, but then accidentally changes an old flow that was already working. Or it cleans up something that looks weird, but was actually there because of a real production bug before. So my workflow became less like: build this feature and more like: read these files first, don t touch ... ... babysit the model. Now that raw coding speed is solved, the next frontier is working creatively away from the desk finishing tasks on a phone (or any new device) in seconds. I m tackling that with LLMonster: a mobile