Comment by weitendorf

It's usually not 24/7, it's that when I use just 1 agentic coding tool there's too much variance/hand-holding to go do something else, and also often a lot of downtime in between turns as they work.

So it's a throughput problem. If there are multiple separable-enough things I also want to do, and I'm already sitting there working, I start adding more things simultaneously until I'm fully occupied. I'm also trying to get better at firing off work that is useful enough to be worth doing if it costs $2 and 10m prompting with a 50% success rate, but not worth doing if it costs me 4hr of focused work with a success rate still not close to 100%. I think right now there is a lot to be gained looking at the economics of "more annoying than hard"/grindy dev work. Now you can just pay something purpose-built for grindy dev work a fraction of what it would cost you.

For example, it's relatively common for me to need to build Envoy with bazel and wait 1-2hr for my LLM agent to get past that (because it maybe takes 30m but if it ignores the cache, or makes mistakes requiring do-overs, you multiply that), or I want to have it do a bunch of benchmarking or testing that has a 10m-20m iteration time and a lot of things to iterate on. I just let them have at it and check to make sure they don't get stuck, and do something else.

I'm also trying to get better at firing off work that is useful enough to be worth doing if it costs $2 and 10m prompting with a 50% success rate, but not worth doing if it costs me 4hr of focused work with a success rate still not close to 100%, because I think right now there is a lot to be gained looking at the economics of "more annoying than hard"/grindy dev work now that you can just pay something purpose-built for grindy dev work a fraction of what it would cost you, or that can read more words in a minute than you can in one day. If it doesn't work or I just abandon it halfway it doesn't matter because it cost me $1 and if it's important I'll just pay one more $1 later.

I had three subagents working for 15m each earlier today gathering a bunch of data about a software dependency I was evaluating using, and they went through a ton of git commits/open source committee meetings/roadmap and historical data, and reviewed all its code. At API rate maybe it would have cost >$10, but I pay subscription rates and it was basically just a 5m writeup of resources to check and things to clarify/evaluate for me while I waited on another thing. To review all that output, and coordinate it/save it, and make it easier to include artifacts, code, etc. I just tell them to write it up in a git repo and push when they're done.

IMO the biggest bottleneck with agentic development is the iteration-time and how much human effort it takes to do the last 20% to take it from slop to actually-done, IMO. But I think many things become worth doing to a slop-level of quality if the time/compute is cheap, it's useful enough to you, and the primary barrier to having it is just taking the time to ask for it.