初创公司 Rivet(YC F25)发布了一款针对大规模 AI Agent 工作负载的编排工具 1。该工具采用 Actor 模型而非重量级沙箱来运行 AI Agent,并利用 S3 进行持久化存储 1。其核心架构通过 SSD 缓存优化读写性能,旨在为大规模 Agent 工作负载提供高效、低成本的基础设施解决方案 1。
在资源效率方面,Rivet 运行的 Actor 基线内存占用仅为 72 KB,相比之下最小沙箱方案需消耗 512 MB 1。在启动速度上,Actor 的 p99 启动时间为 4.9 ms,而沙箱方案的 p99 启动时间为 1.7 s 1。此外,数据存储于 Rivet 上的成本比在 EBS gp3 上便宜 10 倍 1。在三个可用区配置下,其持久化写入的 p99 延迟为 1.07 ms,读取的 p99 延迟为 1.36 ms 1。
在实际应用部署中,客户 Amp 已使用 Rivet 编排多人 Agent,并在周末从 Cloudflare Durable Objects 迁移至 Rivet 1。目前,Rivet 支持多种部署方式,包括单 Rust 二进制文件自托管、Rivet Cloud (AWS/GCP) 及 BYOC 1。
Rivet, a Y Combinator Fall 2025 startup, has launched an AI agent orchestration tool built on the Actor model rather than traditional heavy-weight sandboxes 1. By leveraging Amazon S3 for persistent storage, the platform aims to provide high-density infrastructure capable of handling large-scale agent workloads with significantly reduced resource consumption 1. The architecture utilizes SSD caching to optimize read and write performance 1.
The technical specifications highlight a stark contrast between Rivet’s approach and conventional sandbox methods. A baseline actor instance occupies just 72 KB of memory, whereas the minimum requirement for a standard sandbox is 512 MB 1. Furthermore, the p99 launch time for actors is recorded at 4.9 ms, compared to 1.7 seconds for sandboxes 1. Data storage costs on Rivet are reported to be ten times lower than those on EBS gp3 volumes 1. Performance benchmarks indicate durable write latencies of 1.07 ms and read latencies of 1.36 ms across three availability zones 1.
Rivet offers flexible deployment options, including self-hosting via a single Rust binary, managed hosting through Rivet Cloud on AWS or GCP, and Bring Your Own Cloud (BYOC) configurations 1. Early adoption is demonstrated by Amp, which migrated from Cloudflare Durable Objects to Rivet over a weekend to orchestrate multi-user agents 1.
评论
还没有评论,欢迎留下第一条。