When agents build, deploy, and maintain, persistence becomes the hard problem

Every app has always needed somewhere to keep its state. What’s different now is who creates that state, how many copies get created, and how long they have to hang around after nobody is watching anymore.

The clearest example here is Kimi, Moonshot AI’s platform. A non-technical user describes what they want in plain language, and the agent builds the frontend, backend, and database, then deploys everything in minutes. But it doesn’t stop there. It comes back later to maintain the app, which is where the old assumptions start to crack.

That’s because agents create far more casually than humans do, and most of what they create goes idle almost immediately. A per-tenant database instance looks tidy on paper, but at tens of thousands of tenants it becomes expensive in exactly the wrong way: cost rises with what has been created, not with what is actually being used. Kimi’s earlier setup used single-instance PostgreSQL, and that kind of design runs into the idle cost trap once the tenant count gets high enough.

The fix is to separate durable state from ephemeral compute. Kimi does that with a virtual layer over shared storage, plus a routing layer that sends requests to the right place. For a new tenant, provisioning takes about a second from a warm pool of pre-initialized resources. That matters because if the system is slow to spin up, you have just traded one problem for another.

The same logic applies to the agent’s own workspace. If an agent has to return days or weeks later, it can’t lose source code, Git history, checkpoints, or task progress every time the environment is torn down. Kimi keeps that development state in a persistent filesystem, so the agent resumes instead of rebuilding its own context from scratch. And that consistency appears to help the output too: Kimi saw code generation success rates improve after standardizing on a unified data layer.

The bigger point is that agent-era infrastructure has four jobs at once: isolate tenants, provision instantly, keep idle cost near zero, and preserve state independent of compute. Each of those has existed before. What’s new is having to do all four at once, across tens of millions of agent-created tenants and workspaces. That’s the real pressure point now.