Rivet open-sources actor-based agent orchestrator built on S3

Rivet, a YC F25 company, open-sourced an actor-model orchestrator for agent workloads under Apache 2.0, reporting 4.99 ms p99 cold starts and a 72 KB memory baseline per Actor. The storage layer is SQLite databases backed by S3, with the team claiming 1.07 ms durable writes and 1.36 ms reads at p99.

Each Actor is "a long-lived process for long agent turns, in-memory state for context, a WebSocket server for multiplayer, durable SQLite storage backed by S3, and hibernation when idle," according to the Launch HN post from founders Nathan Flurry and Nicholas, who previously built real-time infrastructure for the Diep and Krunker multiplayer games. They migrated Rivet's storage from FoundationDB to an S3-backed design that acknowledges writes once SSDs in two of three availability zones confirm, then batches to S3 asynchronously, which they say runs 10.4x cheaper than replicated SSD storage.

The pitch is a frontal shot at Cloudflare Durable Objects. Per the GitHub README, Rivet drops the 128 MB memory cap and the 10 GB SQLite limit, removes random Actor evictions, and runs on vanilla Node.js or Bun instead of a custom runtime. Flurry wrote on X that Rivet is "39x faster" at actor creation, "28x faster" at SQLite writes, and "13x faster" at requests from the backend than Durable Objects.

The numbers are all self-published; neither the HN post nor the README names the test workload, hardware, or dataset behind them. The project cites 6.7k GitHub stars at launch and lists one named migration, agent startup Amp, which the founders say moved on-prem over a weekend. Agent infrastructure has been a dense beat for us lately, with 538 agent stories tracked over the past 90 days.