Hey @Luca, thanks for this great article (and short - thank you for that - lots of value in 5mn reading is very pleasant !).
You still run your coding agents on your Mac Mini right ?
Do you know of any "simple" stack to - hopefully at some point - easily create a fleet of agents in the cloud ? I have read a bit about Crabbox (from Peter Steinberger), Rivet which also seems promising - but haven't had a chance to set things up yet. I am looking for assembling a bunch of tools to be able to quickly and easily say : "create an agent with these instructions, and here is the cloud based server you can run it into" (like starting with a cheap 5 euros / month VPS, upgrading as it needs to scale)
Maybe it is a question that fits the Community forum better; in that case let me know I will move it there !
I know that you're moving your AI dev to the cloud, however have you had issues with your local agents crashing your mac mini whilst using Codex?
I actually use Claude Code Desktop and run agents/sub-agents. I find that after a few agents spawn maybe 2 - 8 depending on their tasks, they're able to sink my machine. They completely exhaust my machine of all available resources until the memory-pressure goes through the roof and the swap grows uncontrollably suffocating the machine into submission!
I'm wondering if you've encountered the same and how you've navigated around this?
If anyone else has overcome or even experienced this, I'd be happy to hear about this too. I'm currently trying to work on a skill alongside my project that pre-empts load and causes the orchestrating agents to only act and/or spawn subs when it's safe to do so.
Yes this happens to me! I can’t reliably spawn more than ~2 agents at the same time, otherwise my machine hogs in some cases. It’s a real problem, and I don’t have a good solution: I usually have Codex work on one thing at a time (but 24/7). But anyway it’s one of the many reasons why I believe all this stuff will just move to the cloud
Moving orchestration off the Mac changes the failure mode more than it removes it: in production, the scarce resource becomes shared state and observability. When I ship agentic workflows, I give each worker a hard time and token budget and require it to hand off a small artifact, such as a diff, test report, or decision log, instead of carrying the whole conversation; retries stay cheap and a bad branch cannot burn the budget. The awkward tradeoff is that this is closer to a CI scheduler than a chat session, so debugging needs trace IDs and replayable inputs. Would you optimize the cloud version around durable artifacts or around a long-lived orchestrator context?
Smart shift specialized agents beat one monolithic orchestrator. The use computer less framing is the real metric for effective delegation.
Hey @Luca, thanks for this great article (and short - thank you for that - lots of value in 5mn reading is very pleasant !).
You still run your coding agents on your Mac Mini right ?
Do you know of any "simple" stack to - hopefully at some point - easily create a fleet of agents in the cloud ? I have read a bit about Crabbox (from Peter Steinberger), Rivet which also seems promising - but haven't had a chance to set things up yet. I am looking for assembling a bunch of tools to be able to quickly and easily say : "create an agent with these instructions, and here is the cloud based server you can run it into" (like starting with a cheap 5 euros / month VPS, upgrading as it needs to scale)
Maybe it is a question that fits the Community forum better; in that case let me know I will move it there !
Thanks in advance 🙏
I am working on it right now! Don’t have a good answer yet. I definitely want to move the coding agents to the cloud.
I feel asking this in the community is also good – someone might have done it already!
Hi Luca,
I know that you're moving your AI dev to the cloud, however have you had issues with your local agents crashing your mac mini whilst using Codex?
I actually use Claude Code Desktop and run agents/sub-agents. I find that after a few agents spawn maybe 2 - 8 depending on their tasks, they're able to sink my machine. They completely exhaust my machine of all available resources until the memory-pressure goes through the roof and the swap grows uncontrollably suffocating the machine into submission!
I'm wondering if you've encountered the same and how you've navigated around this?
If anyone else has overcome or even experienced this, I'd be happy to hear about this too. I'm currently trying to work on a skill alongside my project that pre-empts load and causes the orchestrating agents to only act and/or spawn subs when it's safe to do so.
It's been an uphill battle!!
Kind regards,
Daron
Yes this happens to me! I can’t reliably spawn more than ~2 agents at the same time, otherwise my machine hogs in some cases. It’s a real problem, and I don’t have a good solution: I usually have Codex work on one thing at a time (but 24/7). But anyway it’s one of the many reasons why I believe all this stuff will just move to the cloud
Moving orchestration off the Mac changes the failure mode more than it removes it: in production, the scarce resource becomes shared state and observability. When I ship agentic workflows, I give each worker a hard time and token budget and require it to hand off a small artifact, such as a diff, test report, or decision log, instead of carrying the whole conversation; retries stay cheap and a bad branch cannot burn the budget. The awkward tradeoff is that this is closer to a CI scheduler than a chat session, so debugging needs trace IDs and replayable inputs. Would you optimize the cloud version around durable artifacts or around a long-lived orchestrator context?
Having a 'Chief of Staff' agent named Brian who clears your Rome calendar is the most Silicon Valley flex I've seen all week.