If you know me you already know that for the past few months I've been down a major rabbit hole with this idea of cloud agents, agents in sandboxes, that sort of thing. It mainly started with this repository: https://github.com/ColeMurray/background-agents which I have also personally forked in the past in various ways and also contributed to various GitHub issues on.

There absolutely are very, very real benefits to using something like that repository or just otherwise having cloud agents.

As a company offering this sort of thing to your employees allows you to have full observability into all of the sessions. This means you can see how they've prompted the agent, you can see what skills they've used or haven't used, you can see token spend, and you can analyze all of this data to try and improve the company-wide harness.

That benefit in and of itself is tremendous, but there are a few more.

You can, through what you learn over time, systematize and mandate a certain process, a certain SDLC. You want working on a linear issue to start with a certain skill? You can force it. You want that whenever a PR opens, a certain risk assessment agent absolutely always runs. Cool, you don't have to count on your employee's laptop or your employee to do that. You just have a PR-triggered agent session start with the specific instructions.

Maybe whenever a Sentry alert fires up, you want the first responder agent to take a look at the logs, traces, metrics and any recent PRs in the past N hours, and automatically diagnose. Maybe even roll back if needed so that by the time your on call developer gets out of bed and at his desk to take a look at the thing, there's already been some work done about the incident.

You simply want to use multiple types of models from multiple types of providers and just generally be able to flexibly and easily cross-company move between providers? A cloud agent harness like the one I linked to above has your back.

Do you just need to run, for whatever reason, a ton of agents concurrently? You can scale very high when your compute is in the cloud instead of the compute being your employees' laptops.

Hell, maybe you don't want to buy expensive laptops for your employees because your turnover is such that it ends up being cheaper to rent/buy hardware in a data center. Awesome! Cloud agents galore for you.

These are all amazing benefits but they do not come without a cost.

I've been using a fork of background-agents for 2 months and a half now and I've been using it daily. Mostly on my own projects.

The one big thing that I've learned is that these sandboxes are inherently stateless and ephemeral. They die. They just fucking die. They die all the fucking time. That's what they do.

I can't stress this enough. Right now as I am typing this, I have a Codex session which has been running for a total grand amount of 6.5 hours. That is something that absolutely never happened with my cloud agents. They would generally either die within 2 hours or, because you don't really have access to something like goal setting, they would just stop working. These are the very same models with the very same skills and harness and so on.

All of this has meant that I have had to babysit them, checking often that they are still working and that they are still running, making sure to keep nudging them so that they are online. But there's more: there have been various bugs that I have to patch myself with regards to the reliability of these things. All of that has been very time-consuming.

And because this was open source I also had a bit of a nesting effect, meaning that I would often find myself thinking, "It would be really nice if the harness had this feature. Well let me just build it, why not? It's not like I don't have the tokens.” and then I would end up spending even more real time (and tokens FWIW) on building these things.

If you're a big company you can probably afford SREs and full-time developers working on either your fork of something like background-agents or just entirely your own internal harness. You can afford it because the cost gets amortized over a large amount of users.

But I don't work in big companies. I'm a freelancer/boutique dev shop working with early-stage startups with anywhere between 1 and 50 employees.

This means that the amount of time I and my team mates have in a day is limited. It is best that we focus as much of it as possible on the core competencies in the specific project or company we’re working on at that time. This usually means outward-facing product work, not internal dev tooling work.

On top of that I was spending something like $200 every month on sandboxes, which initially didn't feel like much but I also initially hoped I could optimize it further down. Spoiler: I couldn't and, yes, it was even worse initially.

Eventually I had to realize that this is the cost of an OpenAI subscription and also that my business costs, in between all of the AI subscriptions and this, have grown significantly more than makes sense. $1700/mo to be precise (this includes accounting, multiple AI subs, sandboxes, and a few other things).

It's not that I can't afford it. I can but what am I really getting for it? All of the benefits that I mentioned earlier don't really help me with the types of companies that I work with. Even the first responder through Sentry incidents. It's cool, sure, but let's be real: in a company where you have like 6 developers, nobody genuinely expects a 10-minute incident response time unless it’s a hack (and even then, at 10 minutes you’re probably too late).

So after all of this what I want to say to everyone that I have been peddling cloud agents to: I was wrong. They can make sense, but they don’t necessarily make sense.

I actually tweeted this whole rant in shorter form earlier today and somebody told me to try a VPS instead of using these ephemeral sandboxes. To that I say, fuck no.

Yes it may fix some of the problems that I've been fighting with but ultimately it still requires you to set up a new custom system in some sort of way. Whereas I can just run Claude Code or Codex on my laptop and just keep the laptop open 24/7, which I honestly already was doing. I don't have to think about a new setup or a customized setup. It just works.

Maybe if someday cloud compute becomes more accessible, maybe as the tokens themselves become more accessible, who knows? We might see a wave of cloud agents but for me, unless somebody figures out a way to offer them cheaply and with very low friction: no thanks.

This closing note does give me flashbacks to something I said a couple of newsletters back:

P.S.: another interesting benefit of cloud agents I just realized is if you're a digital nomad and you just can't keep your laptop open 24/7, then yes, cloud agents would help you.