Each individual task is like 1.5x slower when using AI But you can have multiple agents working on multiple tasks at a time The engineers who learn to get good at microing +3 agents concurrently will see massive productivity boosts

Dillon Mulroy λ @dillon_mulroy

4 weeks ago

Each individual task is like 1.5x slower when using AI But you can have multiple agents working on multiple tasks at a time The engineers who learn to get good at microing +3 agents concurrently will see massive productivity boosts

102 180 3K 121K 255

@zack_overflow I get the sentiment but isnt that .5x extra slowdown because you end up having to go over what the agent did etc? Like this is out of genuine curiosity. Once you get to 3+ agents do you think the cumulative time you’re spending fixing their stuff is worth it?

@zack_overflow gotta work on that apm

@zack_overflow i shouldve played starcraft as a kid

I think that's true for realms where programming is not theory building, but rather cranking out code. IMO, you can't theory build in parallel like that. And if your job is cranking out code, you're NGMI? Because those jobs are going away. There are plenty of situations where you use programming as a tool to get something done, where you need to crank out code. And AI will be incredibly useful there. But I think that the total human time spent "cranking out code and not thinking too hard about it" will drop, so if that's the core aspect of your job, you're in trouble.

@zack_overflow Or do one task yourself and have the agent hammer on a few other things in the meantime

@zack_overflow its like hiring 3 interns who learnt to code by books and can just try stuff and get it wrong in the real world and iterate without input, if they are wasting time then fine you are paying for it.

@zack_overflow in the next version of Bun?!

@zack_overflow Bro please. I just spend half a day dealing with aftermath of missing Claude code forgetting that we’re running MySQL instead of Postgres and using returning clause a lot. I couldn’t possibly handle 3 times that

@zack_overflow nah, skill issue

But the hard part of dev isn’t cranking out a code, it’s deciding: Which tradeoff is less painful in the long run? How to design so it scales tomorrow, not just works today? Where to reason through edge cases nobody even thought of? And if you’re waiting on AI to do it for you? You’ll waste more time fixing its mess. Better to think it through yourself first

@zack_overflow When I first started using Claude code, yes. After a lot of work, definitely no. The only limit is how fast I can read the code, or at least glance

@zack_overflow It’s also not too hard to bring that 1.5x down to like 1.2x or even closer to 1 with a full pipeline. Then you can sometimes get multiple tasks in the pipeline, think cpu pipelining, and get better throughput.

@zack_overflow Few get this

This is exactly it. I see people trying to replace their entire workflow with AI and getting frustrated. The sweet spot is using AI for the grunt work while you focus on architecture and decisions. Let Claude handle the boilerplate, you handle the brain stuff. That's when you actually 3x your output.

@zack_overflow Sounds like a product manager to me

@zack_overflow so basically, you're saying engineers need to become AI multitasking ninjas. imagine the chaos when they start sending their agents to do the dishes too. productivity or madness? you decide.

@zack_overflow we have now reached "oh ai makes you more productive if you have 1000 hours on starcraft"