Rivet Brings Actor-Based Orchestration to Agent Workloads with S3-Backed Storage
Rivet applies actor model orchestration to AI agents, pairing lightweight actors with tiered SSD and S3 storage to reduce idle sandbox costs and cold start latency.
From Game Infrastructure to Agent Orchestration
Rivet began as infrastructure for multiplayer games such as Diep and Krunker. Co-founders Nathan and Nicholas, who went through Y Combinator in 2023, spent time optimizing high-throughput matchmaking and real-time player messaging. When Postgres couldn’t keep up, they first assembled Lua scripts on Redis, but maintenance became difficult. The team extended its game server orchestrator into what is now Rivet Actors, using many lightweight workers with in-memory state and durable storage. According to Rivet, that system served tens of millions of monthly active users.
The shift to AI agents exposed the cost of running a sandbox per agent. Rivet’s comparison puts actor baseline memory at 72 KB against 512 MB for a minimal sandbox, and p99 cold start at 4.9 ms versus 1.7 seconds. The actor numbers come from Rivet 3.0.3 benchmarks; sandbox figures reference E2B. The company argues that agents spend most of their time waiting on inference, and sandboxes remain idle.
Tiering S3 and SSD for SQLite
Rivet’s storage design keeps S3 out of the write path. Writes to an actor’s SQLite database are acknowledged once SSDs in two of three availability zones have the data, then batched to S3 in the background. Reads check SSD cache first and fall back to object storage when needed. Published p99 latencies include 1.07 ms for durable writes across three AZs, 1.36 ms for reads, and 138 ms for reads from object storage. Rivet says storage costs are about 10 times lower than EBS gp3.
The company previously ran on FoundationDB. Customers stored far more data in actors than in Postgres—session transcripts, agent artifacts, full POSIX filesystems—and SSDs often sat idle holding old transcripts. Rivet now uses S3 as primary storage and treats SSD as a tunable cache tier. That is the architecture behind the project’s “S3 at SSD speeds” claim.
Deployment and Early Adoption
Rivet ships as a single Rust binary with S3 as its only required dependency. The same control plane runs in local development, self-hosted environments, Rivet Cloud, and BYOC deployments. Rivet Cloud provides dedicated clusters on AWS or GCP in any region; BYOC places the control plane inside a customer’s VPC while Rivet operates it. The project is open source under Apache 2.0, and actors can be written in TypeScript or Rust. It also lists React hooks, cron scheduling, actor-to-actor calls, scoped JWTs, and OpenTelemetry support.
One named customer is Amp, which builds multiplayer agents. Amp initially tried Cloudflare Durable Objects but could not launch due to platform limitations and support issues, then migrated to Rivet Actors over a weekend and deployed on-premises. Rivet positions itself as an open-source alternative to Durable Objects, removing the 128 MB memory cap and 10 GB SQLite limit, avoiding random evictions, and running on native Node.js or Bun. Whether these properties hold across diverse production loads is an open question, but the project gives teams another option in the growing category of actor-based infrastructure.