知Yizhi Daily
← Back to this edition

Conversations

Rivet launches S3- and Actor-model-based agent orchestration platform

Edition 2026-11-06 · 08:00 Beijing time

The Rivet team introduced its open-source, self-hostable orchestrator for agent workloads on Hacker News. The platform replaces one heavyweight sandbox per agent with lightweight Actors and uses S3 together with SSD caching for persistent storage, emphasizing high density, fast startup, and low cost.

Team members Nathan and Nicholas say Rivet Actors can provide long-running tasks with resident processes, in-memory context, WebSocket services, persistent SQLite storage backed by S3, and idle sleep. Its model is similar to Cloudflare Durable Objects, but it is open source and self-hostable.

The team previously built matchmaking, real-time messaging, and rapid scaling infrastructure for multiplayer games, serving tens of millions of monthly active users. After shifting to the agent scenario, they concluded that sandboxes are costly, idle most of the time, and slow to start. The comparison provided in the materials gives an Actor baseline memory of 72 KB versus 512 MB for the smallest sandbox, with p99 startup times of 4.9 milliseconds and 1.7 seconds, respectively.

For storage, Rivet previously used FoundationDB but switched to S3 after users stored large amounts of session history, agent artifacts, and complete POSIX filesystems in Actors. When writing an Actor’s SQLite data, the system confirms writes on SSDs in 2 of 3 availability zones and then performs batched background writes to S3, keeping S3 out of the write path. When reading or waking an Actor, it first reads from the SSD cache and falls back to S3 on a miss. The team says this preserves SSD speed while making storage costs 10 times lower than EBS gp3.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader