How many Claude Codes do you run in parallel?
by•
A couple weeks ago, Boris Cherny (the creator of Claude Code) shared a bunch of really useful tips on getting the most out of Claude Code. #1 at the top of the list: do more in parallel. He himself runs 10-15 Claude codes in parallel.
His advice and practice makes sense: coding agents give us the ability scale infinitely. At this point, the only real limiter is our own ability to manage all of these agents.
How many Claude Codes do you run in parallel? And what are your tips for setting things up to maximize this?
651 views
Replies
Tonkotsu
@abedbuilds do you prefer the task segmentation approach because it results in better quality, or because it helps you scale better with less mental load? I personally find that adding structure (first you plan, then you implement, then you validate) helps reduce mental load and improve scale.
@derekattonkotsu Mostly mental load, to be honest. Quality improved as a side effect.
When I ran too many in parallel, I spent more time re-contextualizing than actually building. Segmenting by phase made it easier to think clearly and catch issues earlier.
The plan → implement → validate loop feels more sustainable for me.
I haven't closed the loop fully yet, so I'm still the one to test the feature and look at the UI before it goes online. Therefore, I'm currently maxing out at my brains capacity of focusing on max. 2 different tasks at the same time. (Of course, with exceptions of small tasks. I'm only talking about larger features.)
Tonkotsu
@jorisdejong Yeah that resonates. I'm the same way - I need to hold the quality bar myself if I'm going to put my name to a PR (even if CC wrote it).
In terms of mental capacity, curious what you find more challenging - having to hold all the granular details of multiple projects in your brain at once, or having to constantly attend to various agents when they ask for approval/feedback/etc?
@derekattonkotsu I prefer to work on a single task at a time. I don't think the human brain is designed to do multiple things consciously at once, so I guess I just lean into that. I think a good example is how our eyes focus. You can look at your hand or the background behind your hand, but not at the same time. You can also have a wide focus, but then you're not really focussing on anything.
I know it's a life-long debate, but I, for example, have only browser tabs open related to my current task. I close all the other tabs, because I personally have never really felt like I saved time by keeping a tab open just in case.
How do you see it?
Tonkotsu
@jorisdejong I used to be an engineering manager so I pattern match a lot of this to teams and management. The way to scale as a manager is through parallelization across your team. But a common mistake is micromanagement, which leads to being overloaded because you just can't manage the details of everyone's work - your brain will melt. And I think that's what tends to happen right now with agents.
@derekattonkotsu I agree and that's a nice way to think about it. I'm not sure if I wanna become a manager yet though. I kinda like being a senior developer for now. Just coding :p
UXDesigner.top
I already burn all my credits using 1 or 2 agents almost every day... I guess you need a founder-type subscription to work with 10-15 in parallel! 🤣
Tonkotsu
@david_martin_suarez yeah this seems like a really common pain point. Have you tried the various hooks/scripts that autoresume a session when your usage limit resets? There's also a bunch that poke CC early in the morning to get you an additional 5h window.
UXDesigner.top
@derekattonkotsu Wow!!! I hadn’t even thought about that, but it sounds like an incredible idea. I’ll look into it, especially now that I’m burning through a lot of credits with a new side project.
Sketchize
Just one is more than enough for me right now, for my own sanity. 😅
agents aren't the bottleneck. I am.
vibecoder.date
One.
But I have it spawn subagent swarms for various purposes.
My favorite is the "Snooty Squad"
A set of pedantic and insufferable code reviewers, jaded expert personas from various fields.
The personal/roleplay no joke gives them a 2x improvement in effectiveness.
I stick to 2 instances right now! I find splitting the work between a "main task" and a "support/debugging" one is the sweet spot for me—any more and I start losing track of the context.
Tonkotsu
Interesting to see the responses are really clustered around the low end (at this time, over half of respondents are using 1-2 agents only). It shows the tremendous opportunity ahead of the industry if we can actually put a full team (say 5-10) agents to work for most developers.
For me (and for others judging by the thread), the main blocker is cognitive load. It's draining if you have to constantly round-robin through your agents, answer questions/give approvals in realtime, and remember all the granular details of each agent session. Over time, I've developed a structured approach that I've found helpful:
Limit interactions with agents to two high-leverage moments: planning (so you are aligned from the start and have an opportunity to think through the thorny issues), and code review/verification.
Structure those two moments around clear artifacts. For planning, this means a succinct multipoint plan with key decisions called out. Importantly, it should not be a Markdown novel!
This is essentially what we're building with @Tonkotsu to add structure (and sanity) into parallel agents.