zlacker

>>laiysb+(OP)
Are people really doing coding with agents through PRs? This has to be a huge waste of resources.

It is normal to preempt things like this when working with agents. That is easy to do in real time, but it must be difficult to see what the agent is attempting when they publish made up bullshit in a PR.

It seems very common for an agent to cheat and brute force solutions to get around a non-trivial issue. In my experience, its also common for agents to get stuck in loops of reasoning in these scenarios. I imagine it would be incredibly annoying to try to interpret a PR after an agent went down a rabbit hole.

>>ethano+a5
Googles jules does the same (but was only published yesterday or so). I think it might be a good workflow if the agent is good enough. Copilot seems not to be in these examples and then I imagine it becomes quite tedious to have a PR for every iteration with the AI.