Rivet launches agentOS beta with 6.1 ms cold starts and 32x lower cost

• Rivet launched agentOS in beta as an open-source runtime for AI agents that runs on WASM and V8 isolates, with agents embedded directly into an app process instead of separate VM- or container-based sandboxes, according to Rivet launch and the architecture thread.
• Rivet says the performance delta is large enough to matter for agent infrastructure: its benchmark post reports 6.1 ms p99 cold starts, about 131 MB RAM per coding agent, and $0.00000058 per execution-second on AWS on-demand pricing.
• The packaging is deliberately minimal: Rivet's install post says agentOS ships as npm install rivetkit, keeps "your APIs, your auth, your toolchains," and the portability post says the same package runs from laptops to Lambda and Kubernetes.
• Rivet is also bundling agent primitives that teams usually assemble themselves: the filesystem post describes storage mounts for S3 and SQLite with host-side credential scoping, while the orchestration post adds webhooks, retries, queues, delegation, and persistent SQLite state.

Rivet is positioning agentOS as an in-process runtime, not a hosted sandbox service. In its architecture thread, the company says traditional sandboxes boot full VMs or containers, while agentOS runs agents in lightweight VMs inside the host process using the same isolate model it describes as the technology behind Chrome. The practical pitch is lower startup and less infrastructure overhead, with the install post stressing "No separate infra. No sidecar" and a TypeScript-first setup.

The company’s main performance claim is in the benchmark post, which lists 6.1 ms p99 cold starts versus 3,150 ms for the fastest compared sandbox, plus about 8x lower memory use and 32x lower execution cost. Rivet also published benchmark methodology, which matters because these are vendor-reported numbers rather than an independent bakeoff. The broader implementation point is that if the measurements hold up in production, the runtime starts to look less like remote sandbox provisioning and more like a library-level dependency for agent execution.

A big part of that design is portability. The portability post says the same npm package can run on Vercel, ECS, Cloud Run, Lambda, Kubernetes, or even a Raspberry Pi, and the GitHub repo frames agentOS as open source rather than a managed-only platform.

Rivet is bundling several pieces that teams usually bolt together around an agent runtime. Its orchestration post lists webhooks, "durable workflows with retries," queues for backpressure, agent-to-agent delegation, multiplayer observation, and persistent SQLite state. That makes the beta as much about operating long-running agent workflows as about shaving milliseconds off startup.

The filesystem model is another opinionated choice. In the storage post, Rivet says "Agents think in files" and can mount S3 buckets, SQLite databases, Google Drive, and the host filesystem as a directory tree. The host keeps credential scoping, so agents read and write files without directly handling secrets or API keys.

Security controls are similarly host-centric. The security post says developers can set per-agent CPU and memory limits, programmatic network allowlists, isolated network namespaces, and bring their own auth with API keys, OAuth, or JWTs. Taken together, the beta is less a single-model agent framework than a portable execution layer for coding agents and backend automations that need local tools, state, and tighter control over runtime boundaries.