If your whole workflow is one Claude Code instance on your MacBook, you're leaving 10x on the table. Here's why. On your laptop, running 2 agents at once means two copies of your code, and you're constantly fighting over which worktree you're on and whether everything's synced. It caps... out fast. Cloud agents in VMs let you run as many as you want without worrying about any of that. Each one gets its own machine, so the work never collides. Spin up 10, let them all run, and nothing steps on anything else. Watch full ep below on The Startup Ideas Podcast (SIP) 🧃. Thanks to Ryan Carson for sharing the sauce. Cloud agents are probably the future. So is doing half your work from your phone, which sounds insane until you try it once. Then you don't go back.show more

Respectfully, you clearly don’t understand how git worktrees work if you’re if you’re have contention issues with agents across worktrees.

What’s the difference between this idea and running multiple agents inside 10 different branches inside Claude Code simultaneously? I run multiple agents all day.

Well, actually, you can have 10 sessions run on Claude simultaneously in parallel each of them with their own directory, i.e. workspace. Zero collision.

OR you can run multiple containers in your laptop too, skip the cloud VMs infra. With one dockerfile and docker-compose, you can clone, make changes, commit, push to PR, test, merge, and repeat. OR Tell your agents to use a specific folder /wt for all your worktrees, and use a cleanup instruction after they're done.

True, and the bottleneck moves fast once you do it. Running many agents is easy. Reviewing what they all produced is the part that does not parallelize, so the leverage only shows up if you have built a way to verify output faster than you generate it.

FREE Guide to Run Multiple AI Agents on Your Machine at Once

I disagree here. There are plenty of things we want agents to do that require using and access information you would never want hosted in the cloud, managed by a third party. I know the efficiency gains make sense on paper, but the cloud forward strategy is not a one-size-fits-all. It is frankly, not what most customers will need or want with AI. Sensitive data, pii, personal taste, daily habits / trends…. Maybe that’s something you, me, and Ryan are comfortable with in some cases but definitely not the majority of adopters going forward. m2c

parallelism makes shipping cheaper, but it makes judgment more expensive. the real 10x isn’t ten agents—it’s knowing which decisions can run unsupervised.

This is EXACTLY what solid is built for - your own cloud agents with real machines.

Cloud VMs move the bottleneck from worktree conflicts to review bandwidth. Ten parallel agents is easy, reviewing ten parallel diffs at 2am is the actual limit.

Cloud agents really unlock a whole new level of productivity.

@10_X_eng has been doing this since Fall of 2025... he built:

suggest you take a look at this:

totally getting into cloud agents excited to watch this episode thanks Greg!

"each one gets its own machine" is the part people underestimate. it's not just parallelism, it's that you stop context switching between environments entirely. that mental overhead was eating more time than the actual compute limits.

Treating your local laptop as the runtime for parallel AI agents is the same architectural mistake as running production workloads on a single bare-metal server in your closet. Ephemeral cloud VMs are cattle, not pets. If you are not orchestrating isolated, disposable environments for your agents, you are not actually scaling; you are just bottlenecking your own pipeline.

@grok summarize this video in a detailed outline summary

Matches what I've been seeing. The real unlock wasn't a smarter agent, it was giving each one its own lane so they stop stepping on each other's changes.

The bottleneck shifts from generation to adjudication. Four parallel Claude Code streams means four decision queues waiting on you. The 10x only holds if scopes are pre-defined tightly enough that each agent runs without constant check-ins. Most people discover they're the single-threaded component.

This is quality. Thanks @gregisenberg

or just clone the same repo into different folders, run more than 1 pc at home, claude is already on the cloud, a 500 mini pc is all you need.... how do sell VMs 101 is all this is.

How about raspberry PI 5s, always on, cheaper and pays back within a year, tailscale and access from anywhere?

I don't get how people think you can just let AI do a bunch of stuff without review. Like you still have to spend HOURS of reviewing and making sure the stuff isn't cooked. Yea you can have 10x output for simple tasks but for complex stuff nah. This is just hype.

@gregisenberg totally been there. trying to juggle multiple agents on one machine is a nightmare. split it up and save yourself the hassle, trust me.

I'd already been thinking about this, but I wasn't sure if it would work out🤔

Most miss the opposite pattern for ad ops. Parallel agents wreck creative workflows without a lead agent holding brand voice. Ran multi-agent creative iteration on Gethookd's stack, quality collapsed by round 3. Serial beats parallel when the output is taste.

@grok, what could one learn from a man talking about the importance of security and demonstrating copy-pasting production API keys from 1Password to agents?

Cloud agents not colliding with each other is the real unlock there.

The multiplier isn't just more agents, it's giving each one an environment where it can work independently without creating coordination overhead.

the parallel agents point is real but the hidden cost is context switching for the human. running 3 agents means I'm the bottleneck reviewing 3 streams. curious how you keep your own head from thrashing?

That was great. Thank you.

Cloud agents would still have those issues you’ve just moved the environment

cloud vms are fire but gotta pair that with onchain physics for real game state sync

How about branches?