Why we built Cloudroom to run coding agents in the cloud

What is Cloudroom

Cloudroom is a solution for easily running your coding agents in the cloud. A nice GUI, support for your existing coding subscriptions, and most importantly, an open-source solution where you get to keep all of your agent traces.

We are an alternative to the cloud offerings from Cursor, Amp and Codex without the ecosystem lock-in.

The first startup

A few months ago, the team and I were working on a completely different startup named DeepAPI. It was essentially a router for different tools you could give to your agent. But unlike other competitors which started to pop up after we went live, we were actually choosing all of the tools for the users and doing the testing for them.

It was quite successful. The metrics weren’t perfect, but they were certainly good enough to scale.

But something felt off. After going on a 5-day trip into the mountains to change environments and deliberate, we realized we were heading in the wrong direction. At the start, we thought we would be creating cool endpoints and innovating, but what really gained traction were web search and deep research endpoints, which we weren’t enthusiastic about and had no unfair advantage in.

That’s when we decided we had to move on. It would only be a failure if we stuck around for too long and spent another multiple months or years building something we weren’t passionate about.

Searching for an idea

After DeepAPI, we decided to take a little breather. Before rushing ahead and starting work on the next thing, we decided to reflect for a second and figure out what exactly we should pursue and what we should drop.

The biggest help in this stage were Paul Graham essays and voice recordings. We would walk around the city or sit at cafes and reason deeply about new AI inventions, interesting tweets and questions from the essays themselves – all of this while recording, so that we could later transcribe it and feed into LLMs.

While doing all of this, we followed an idea from one of the essays: startups start as side projects

We would pick a side project to work on and release it after a few days of work. Rinse and repeat. However this process proved tiring, and we didn’t feel like we were getting closer to actually building a startup.

In hindsight, the issue was obvious. The essay didn’t account for LLMs. Before, a side project would mean working on something for a few months, perhaps a few weeks if you were really fast. In this case, AI actually made us move TOO FAST, and we weren’t spending enough time to see if any of the little side projects were gaining any traction.

Promoting a side project that would be replaced next week felt pointless.

Immortal Agents

Despite our flawed strategy, we weren’t completely off the base. Our most directionally correct idea, Immortal Agents, was attempting to solve a sub-issue of the problem that Cloudroom solves today: when you lose internet connection, your agents stop running and they don’t restart.

In theory this sounds like a pretty easy problem to solve. Simply detect when the connection is disconnected, find all the affected sessions, and once the internet is back online, send a quick “continue” message to them.

As we learned however, every single harness has completely different retry behavior. One harness would restart on its own after 20 minutes, but another would only try for 10 minutes, not to mention the handling of sessions varies between each harness as well and a plethora of other unobvious issues.

In the end, David managed to figure the problem out, and the project is out as an open-source repo so anyone can use it.

Cloudroom

After releasing Immortal Agents, we knew that we had to actually focus up and choose the next thing. The side projects were nice, but they weren’t the right abstraction in the age of AI.

The problems we decided to solve were agents not running when we turned off our laptops, and agents being bad roommates.

The former should be pretty clear – when you close your laptop, your agents stop working because they no longer have a place to run.

What I mean by the latter, is that the more agents you run locally on your computer, the more issues you will run into. An agent deciding to nuke your computer by opening Docker and running 1000 tests, another agent clicking around your browser to change settings in Vercel (even though you deliberately didn’t give it access in the first place) and many more edge cases which you would never even think of.

So if you’re a serious developer building with AI, your computer can’t keep up anymore, your agents aren’t running all the time when they could be, but you don’t want to hand over all your agent traces and lock yourself into an ecosystem you can’t get out of, try Cloudroom today, this is exactly why we built it.