
Prediction: Sandboxes aren't a winner takes all market because agents will buy compute one tool call at a time. A lot of the sandbox competition looks like a race to the bottom on price and startup time. The assumption seems to be that one provider wins the developer, then gets all their agent's work. I don't think that's how this market will evolve. One agent might need a cheap CPU sandbox to parse a file, a GPU for the next step, and a machine inside the customer's own infrastructure to touch sensitive data. Why would all three run in the same place? Imagine a spot market for agent compute. Before a tool runs, the system finds the best available place to execute it. First apply the hard constraints: which providers do I trust, where is this data allowed to go, what hardware does this step need? Then compare cost, availability and the time it will actually take to finish. The cheapest hourly rate is useless if startup takes longer than the job. And sensitive data doesn't become okay to send somewhere because that provider is cheaper. But what happens after the agent has done the same job a hundred times? It learns the workflow. It should also learn where each step runs best: which environment needs a GPU, which can use a tiny container, which provider keeps failing on this workload, and which work must stay on your own machines. That is where I think self-learning agents and sandbox routing meet. We've already built some of the foundations in @agentrunai: durable execution, learning across runs, and a common computer layer with providers like Daytona and Modal. My bet is there will be plenty of room for sandbox providers. They'll compete for particular workloads, and the agent will make the placement decision. x.com/JoshARosen/status/2121…
Josh Rosen @JoshARosen Nov 15 Article The State of the Agent Sandbox Does every agent really need its own computer, or can much of its work be accomplished inside a far simpler execution environment? Over the past year, agent sandboxes have evolved from relatively







