Rivet Open-Sources an Orchestrator for Long-Lived AI Agent Workloads
Rivet, a startup from the YC W23 batch, has released an open-source orchestrator designed for agent workloads. The launch post appeared on Hacker News on November 5, 2026, and the code is available in the rivet-dev/rivet repository on GitHub. The project builds on the actor model: each agent runs as a long-lived process that keeps its working state in memory, carries its own WebSocket server, and persists data to durable SQLite storage backed by S3.
The architectural premise is direct. Instead of spinning up a fresh sandbox for every interaction, an agent in Rivet stays resident: incoming connections reach it over WebSocket, working state lives in memory while the actor is active, and durable SQLite on S3 preserves it across restarts. The project positions itself as a self-hosted, open-source counterpart to Cloudflare Durable Objects — an architectural reference point for stateful compute, not a competing product being measured against.
The launch figures are the project's own measurements and carry that attribution. Per Rivet, the base memory footprint per actor is 72 KB against 512 MB for a minimal sandbox, and actor startup sits at 4.9 ms p99 against 1.7 seconds for sandboxes. The gap follows from the design — an actor is a lightweight process inside an existing runtime, while a sandbox is a fully isolated environment with its own startup and memory overhead. These numbers describe one project's benchmarks and have not been independently verified.
Third-party usage is already in place. Amp, the coding agent, runs multiplayer agents on Rivet in production — several participants interact with agent processes concurrently, which is the workload class the project targets.
📅 November 5, 2026 • Hacker News
Impact: for teams building agentic systems, the pattern matters as much as the product. When agents must hold state across interactions, keep persistent connections, and recover after crashes, an actor layer with durable storage addresses that directly. The open-source release makes the approach self-hostable, and the published repository lets teams inspect the runtime mechanics — state handling, WebSocket transport, and the SQLite-on-S3 durability path — before committing to an infrastructure decision.
Sources: news.ycombinator.com, github.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)