spion

I sent this to someone, figured I'd post it here too. there's different "buckets" of types of software work: non-urgent bugs: most of the time, you just need to get these fixed, you don't need to babysit it, so you just put the stack/logs/report link into the autonomous overnight queue, and it fixes it when it gets to it random low-confidence experiments: you give a rough sketch, and queue it up as an experiment, and you get the agent's rough conception as a way to evaluate if it's worth investing more time and tokens into or not well-understood features and improvements: you just spec them out using an agent: just kinda guide the agent to a solution / plan, then when you have enough confidence the agent understands what you want, then you queue up the plan in the autonomous overnight queue -- you do NOT finish it yourself, waste of time creative/iterative work: you need to look at stuff as you try it and iterate quickly, so this is the prompt & iterate approach. your agent runs only the fastest/lightest checks, nothing that takes too long. you see the prototypes in front of you as you get things done. when you feel like you've "got" it, and the agent understands it, you then ... put it into the queue and it actually polishes it up and ships it later wide refactoring sweeps like tech debt: you spend some time iteratively fixing a couple of instances to make sure the agent understands what you mean in all types of instances (building up a plan doc as you go), and then queue up a codebase-wide sweep to get it all cleaned up at once the idea is that the overnight queue does all the exhaustive checks and whatnot, and you don't need to worry about it taking 2 hours, because it's not holding you up anymore. It's a constant stream of work being done, like CI, integrating behind you, you're ahead of it, and it's doing the boring long work for you so you focus on creative work mostly, and a few other things, including testing, spot-checking code, improving the validations. it unblocks you entirely from having to wait for agents you really should have 1 or maybe 2 concurrent interactive agent sessions at a time max -- if you are waiting on agents too much, they're doing too many checks or deep work, stuff that could be done autonomously later. just prototype, fast iteration focus you are NOT juggling lots of agents you are NOT getting spread thin you are NOT having to wait for long checks and validations you are spending your valuable time on what humans do better than agents how do you get here where this can be effective? well, I have a TON of thoughts on that ... reach out