Launch HN: Rivet (YC F25) – Scale and Orchestrate Agents on S3

Rivet is an open-source orchestrator for agentic workloads, built around lightweight, keyed “Actors” rather than a separate virtual machine or sandbox for every agent. Each Actor is a long-lived process that can hold in-memory state, use a persistent SQLite database stored on object storage, accept HTTP and WebSocket connections, and run scheduled tasks. Developers can create Actors for agents, users, sessions, or tenants; Actors can also communicate with one another. The design aims to make many independent, durable processes practical while allowing idle ones to hibernate and resume without losing state.

Rivet reports 72 KB of memory per Actor and a 4.99 ms 99th-percentile cold start, compared with 512 MiB and 1.70 seconds for a typical sandbox; it also cites low-millisecond database reads and writes using SQLite backed by S3. These figures come from its published benchmark, while the sandbox figures use documented defaults and a third-party benchmark. The platform includes workflows that can replay after interruption, real-time broadcasting, fault recovery, typed SDKs, observability, and optional AgentOS for securely running agent-written code. It runs with Node.js, Bun, or Rust, and can be self-hosted or deployed through Rivet Cloud or into a customer’s cloud account. Object storage is its only required infrastructure dependency, with support for S3-compatible services and other cloud storage.