A Computer for Every Agent in Software Development

📝 A computer for every agent What will the future of software engineering look like? Maybe we're all "builders". Maybe AI writes all of our code. It's hard to know for sure... but what I find interesting is to look at history. This idea of "build software in the cloud" has been appealing for some time. There's a number of problems we deal with to run software locally: The security model for accessing files/folders (e.g. sandboxing) The fragility of local environments (e.g. Docker, Nix, etc.) The power of the hardware (e.g. consumer hardware vs beefy VMs) The availability of the environment (e.g. laptops going to sleep) And yet... most software still happens on localhost! Why is it that even with all of these challenges, cloud computers haven't won? I've joked that this idea of a "cloud computer in a box" has been launched 100 times. It started back in the day with "dev containers" and "cloud IDEs" to remove the drudge of the four points above. Super fast hardware, always available, immediately accessible for your team. In reality, you only ended up seeing this at companies like Meta. Of course they wouldn't store secrets on their machine in .env files! They would have nice cloud environments they could connect to for everything. But this required a significant investment in local/internal DevEx. Many startups tried to bring this idea to developers more broadly, but I would argue it never really got product market fit. Including GitHub, where you can press "." to open a cloud VM ("Codespaces") with your repo. So... right idea, wrong timing. Fast forward a bit in history, and we started to get the first powerful AI models. All of a sudden there was this new demand for cloud computing. Except the models were still pretty bad! They struggled to use computers like humans. It wasn't until roughly Sonnet 4 that models started to become good at "tool calling". And tool calling is important, because it's a generic abstraction that you can train the models to be good at. You can just give them new tools! In the past year... yeah, the models have gotten much better. So why now, why is this idea of cloud computers finally starting to work? It's the combination of three major things: Better models Better products New capabilities The models can be used as effective agents. They can reliably call tools, and can be given new powerful tools. This includes teaching them to use computers like humans, navigating and clicking around autonomously. The products and agents using the models have evolved to harness this new intelligence. The ideas in the video below simply *did not work well* even 4-5 months ago. The frontier is moving fast. Going back to the local pains from the start, these secure cloud sandboxes do have an improved security model, better hardware, and are always available (your agent can run while you sleep). Software engineering is marked in capability jumps. A computer-using agent and a nice UI is helpful, but not sufficient to change the way we work (changing patterns is hard!). But new capabilities, like agents being able to test their work and send you demo videos proving it's correct, are going to meaningfully change how we build software. It was the difference internally at Cursor from ~8% of PRs merged with cloud agents to ~30% in a few months. That is staggering. Still, even with this new capability, the existing software engineering best practices remain. We're still looking at code. We're still thinking deeply about the system architecture. But it does provide a glimpse into what the future might look like. To me, that future is one with better abstractions. Removing the boring parts of building, and allowing developers to focus on bigger problems. But we have a long way to go, and so much more to build! http://x.com/i/article/20245922496367370…