It's a curse I know, parallel lanes, context switching. We're told to aim for surgical precision, one item at the time. But when the world is keeping endless inspiration, software commoditised it's hard to live by the rule.
Running several agents at the same time, I found myself having a hard time catching up on everything produced, and giving proper feedback. I tend to actually ask my agent to go through transcripts to find old unfinished tasks and re-state the outputs.
MD files became tedious, screen shots didn't show the entire flow, production code was usually stale or on stuck on a branch.
People out there, do you keep to the rules of capping context switch, or what's the best way to review output here?
Coming from reviewing mainly MD files, or text output in the terminals I was positively surprised when Cursor Cloud used its VM to make nice screen shots and recordings so you could see the new feature live before deploying.
But doesn't really solve the option A option B review types
Yes but i manage it as i have another AI that does the code review and another one that does the testing and for all that issues you listed i made my own AI eng team that tracks and manages all of that www.bridge-terminal.com/atlas
Yeah that's the vision for sure. I'll check it out, haven't been able to fully commit to an complete autonomous system
Yeah its difficult, most likely impossible when you add more agents.
I'm building GitComet to help with the reviewing... at least the tool does not slow down and difftastic integration helps to reduce some noise...
But yes there is a lot to review