Rivet 3.0 preview is an open-source runtime for stateful actors, built for teams running agents and other long-lived workloads that each need isolated memory, compute, and a database. The release adds WarpDrive, which gives every actor its own SQLite database. Writes are acknowledged once they land on disk in a majority of three availability zones, then cold data is offloaded to object storage in the background, so S3 is not on the request path. Object storage becomes the only dependency, and Rivet runs as one stateless binary, replacing the Postgres or FoundationDB plus NATS stack required in 2.x. Rivet reports 0.69 ms SQLite writes at p50 and a 6.9 ms cold start when an actor's pages are already on disk, with an actor size of about 72 KB, set against roughly 200 MB for a Kubernetes pod. Idle actors hibernate until the next request, and if a machine fails the actor can move to another node with its state intact. The preview uses the same RivetKit SDK as 2.x, so application code does not have to change. Multi-region, serverless, and ACLs are not in this build, and Rivet says teams that depend on them should stay on 2.x for now. You can embed the binary, self-host it in Docker against Amazon S3, Google Cloud Storage, Azure Blob Storage, Cloudflare R2, or MinIO, run it as bring your own cloud inside your VPC, or use Rivet Cloud. Rivet is Apache 2.0. Founders Nathan Flurry and Nicholas Kissel started the company in Y Combinator's W23 batch, and it is also backed by a16z Speedrun.