Give an Agent a Real Computer and It Stops Being a Coding Tool

There's a running joke on our team that I'm king of using the product for the things it wasn't built for. I gave it access to my Spotify and had it moonlight as a DJ. The team treats it as a bug. I think it's the whole thesis.

Here is the mechanism. We built a coding platform: a full VM in the cloud, real tools, an easy way to talk to the agent, and reviewable output. But once you assemble those pieces, something interesting happens. If the agent doesn't happen to call the tool that opens a pull request, nothing breaks. It just does the non-coding thing instead. Give an agent a real computer and real tools, and it can do almost anything you can describe to it. The coding surface was never the boundary. It was one path through a general execution engine.

This is why I don't wave off the people who repurpose our product for jobs it wasn't scoped for. A prospect this week wanted his ops lead, not an engineer, to use it for everyday tasks. My instinct wasn't to redirect him to a different tool. It was to lean in, because that pull is the agent as the primitive, not the automation. When users keep dragging your product toward work you never designed, they are telling you the primitive is general enough to matter.

The strategic point for anyone building or buying agent infrastructure: the leverage is in the substrate, not the surface. Coding is a good wedge because it forces you to solve the hard problems first, real sandboxing, fast VM startup, provider switching, live terminals. But once those problems are solved, the same substrate runs finance workflows, ops tasks, research, anything a person with a computer could do. That is why agents are about to break out of engineering entirely.

So watch what your users do when they misuse your product. The teams who treat that as noise are optimizing a feature. The teams who treat it as signal are discovering a platform. Build the substrate well enough and the expansion writes itself.