Comment by remram
17 hours ago
Why do you need to "manage" your coding agents like they are people? How long does it take them to do a task once prompted, in the background?
Don't you just prompt an immediately review the result?
17 hours ago
Why do you need to "manage" your coding agents like they are people? How long does it take them to do a task once prompted, in the background?
Don't you just prompt an immediately review the result?
Currently - claude code is pretty slow. I've definitely had it take >15 minutes on the (faster than opus!) sonnet model just thinking and writing code without feedback or running long lasting tools. I expect this to change given that companies like cerebras exist and seem to know how to generate tokens in much less time, but the current state of the art is what it is.
Always - if you're going to pipe the result of some slow process back to them (like building a giant C++ project that takes minutes/hours, or running a huge set of tests...)... it's going to be slow.
For me, the average task takes a coding agent 2-5 minutes to complete. A slightly annoying amount of time as I'm prone to getting distracted while I wait.
This gives me something to do in that time.
My guess is time to complete a task will oscillate - going up as we give agents more complex tasks to work on, and going down with LLM performance improvement.
I don't think smart ones are that fast. They can work for hours if you have the budget.
Interesting, I didn't know that.