Rivet Launches Orchestrator for Agentic Workloads with Low Cold Start Times

Rivet's Core Offering

Rivet, a Y Combinator W23 company, introduced an orchestrator designed for agentic workloads. The system is built on the actor model, where each Rivet Actor functions as a long-lived process. These actors maintain in-memory state, incorporate a SQLite database, and utilize WebSockets for real-time communication. They are addressed by a key and can be created per agent, session, user, or tenant.

Performance and Resource Efficiency

The platform reports performance metrics including 4.99 ms p99 cold start times, attributed to not requiring image pulls or VM boots. Each actor consumes 72 KB of RAM, as actors share a node process rather than each booting a sandbox, allowing a single server to host hundreds of thousands of actors. SQLite databases are backed by S3, achieving 1.07 ms durable writes and 1.36 ms reads at p99, which is approximately 10.4 times less than replicated SSD storage. Actors hibernate when idle, incurring no cost and resuming with state intact upon the next request.

Developer Integration and Use Cases

Rivet supports development in plain Node.js, Bun, or Rust. The provided code examples illustrate how to define an actor with a per-actor SQLite database for persistence across restarts and hibernation. It also shows how to handle chat interactions, durably save messages, stream responses, and broadcast tokens to connected clients. On the client side, developers can connect to an actor, listen for real-time events, and call actions on the actor.

Rivet positions its Actors as foundational components for various applications, including durable processes for agents with memory that survives restarts, multi-step workflows that replay, and an 'agentOS' for running agent-generated code. All products shipped by Rivet use this same control plane and actor scheduling.