How many Claude Codes do you run in parallel?

the engineering manager analogy is spot on. but i think the real unlock isnt running more instances manually - its having one persistent agent that can spawn and coordinate sub-agents on its own. you define what done looks like, it handles the decomposition and scheduling.

i went from juggling 3-4 sessions where i was constantly context switching to basically doing async review of completed work. the difference is massive. instead of being a human router between agents, you become closer to a product manager who checks in on outcomes.

the 10-15 parallel number boris mentioned only makes sense if your tooling handles the coordination layer. otherwise youre just a very expensive load balancer.

i run one persistent agent 24/7 that spawns subagents as needed. so technically "one" but it can kick off 5-6 parallel tasks when i need it to. the trick for me wasnt adding more CC instances, it was investing time in really good AGENTS.md and instruction files so each subagent knows exactly what to do without me babysitting it. once you get the upfront config right, the cognitive load almost disappears because youre not context switching between agents - theyre just doing their thing and reporting back when done. the people running 10-15 manually are probably spending more time managing agents than actually building.

the number doesnt matter nearly as much as how autonomous each one is. i was stuck at 2 for months because every agent needed constant hand-holding - approve this, clarify that, fix the merge conflict. once i started writing proper AGENTS.md files with clear constraints and letting them make decisions without asking, jumped to 5-6 comfortably. the real unlock isnt more terminals, its less interruptions per terminal.

Claude does my tasks quite fast now. If in November it could take up to 30 minutes, similar task now requires 3-5. During this time if I will switch to another project, I will loose focus, so I just could work on backend/frontend/marketing of the same project. Other tasks which can be executed in parallel can be something not related to my programming tasks at all, for example, removing unnecessary files on disk.
Maybe I can have a research task in background, but still people can't handle 15 tasks in background in my opinion

i usually run 2-4 depending on what im doing. the bottleneck isnt the agents though - its me. keeping track of what each one is working on gets exhausting fast.

what helped was being really strict about task scoping before spinning one up. if i cant describe the task in 2 sentences with clear success criteria, its not ready for a parallel agent. otherwise you end up with 5 agents all asking clarifying questions at the same time and youre just context switching between chat windows which defeats the purpose.

the sweet spot for me is one agent on the main feature im building, one on tests, and maybe one doing a refactor or docs task thats fully independent. anything more than that and quality drops because i cant review the output fast enough.

honestly the number of parallel instances matters way less than how well you set each one up before letting it run. i went from constantly babysitting 2 agents to comfortably running 4-5 just by writing better instruction files upfront - like detailed AGENTS.md with project conventions, error handling patterns, what to do when stuck.

the engineering manager analogy someone mentioned above is spot on. a good manager doesnt check in every 5 minutes, they set clear expectations and let the team execute. same thing here - if your agent keeps asking you questions or going off track, thats a setup problem not a scaling problem. i spend maybe 15 min writing instructions per task now and the agents run for 30-40 min without needing me at all.