Rivet is a core infrastructure provider in the AI agent stack, specifically addressing the "compute and state" layer. While most of the ecosystem is focused on the LLM models or the orchestration frameworks, Rivet builds the actual environment where agents live and work. By providing a stateful Actor model, they enable the kind of long-running, multi-turn interactions that modern agents require without the performance overhead of traditional databases or the timeouts of serverless functions.

They are particularly relevant to builders who need to execute agents in production environments where security, latency, and cost are critical. Their agentOS runtime provides a sandboxed environment for tool execution and code generation, which is a necessary component for any autonomous agent that interacts with the real world. By pushing forward the use of V8 isolates and Wasm for agent execution, Rivet is helping define the standard for portable, high-performance agentic infrastructure.

The transition to stateful compute

Most modern cloud infrastructure was built for the request-response cycle of the web. AI agents, however, do not fit this model. They are long-running, maintain complex memory, and require persistent connections to handle multi-step reasoning. Rivet was founded to build the compute primitives that these "thinking" applications require. At the center of their offering is the Actor model, which provides a way to co-locate compute with in-memory state. Unlike a traditional serverless function that must fetch data from a database at the start of every execution, a Rivet Actor keeps its state alive while active and hibernates when idle.

This architecture solves two primary problems for agent developers. First, it eliminates the latency of database round-trips for session context. Second, it handles the complexity of state management across thousands of concurrent agent sessions. These actors can scale to zero to minimize costs and burst to thousands of instances when traffic spikes, making them more economical than traditional always-on servers but more capable than stateless lambdas.

agentOS: The agentic runtime

While Actors handle the state, Rivet's second core product, agentOS, handles the execution. It is a portable, lightweight operating system built on V8 isolates and WebAssembly. This choice of technology is deliberate. By avoiding full virtual machines, Rivet achieves cold start times of roughly 6 milliseconds and a memory footprint of just 22 MB. For developers building agents that need to execute untrusted code or use external tools, this provides a secure sandbox without the performance penalty of traditional containerization.

agentOS is designed to be embedded directly into existing backends. It supports standard runtimes like Node.js, Bun, and Deno, meaning developers do not need to rewrite their stacks to use it. It includes built-in support for tool registries, file system mounting (including S3 and SQLite), and network security controls. This portability allows developers to run the same code on their local laptop, in a private VPC, or on Rivet's managed cloud network.

Origins and market position

Founded by Nathan Flurry and Nicholas Kissel, Rivet emerged from the Y Combinator ecosystem with a focus on solving the infrastructure bottlenecks of the AI era. Based in San Francisco, the team has focused on an open-source-first strategy. The agentOS runtime is Apache 2.0 licensed, allowing for self-hosting on Kubernetes, Railway, or Vercel. Their commercial model follows the standard "open core" playbook: the core runtime is free, while Rivet Cloud provides a managed global edge network with built-in observability tools.

Competitively, Rivet sits between lower-level infrastructure providers and higher-level agent frameworks. While frameworks like LangChain focus on the orchestration of prompts and tools, Rivet focuses on where that logic actually runs. Their platform includes a built-in SQLite viewer, workflow state inspection, and event monitoring, providing a level of visibility into agent execution that is often missing from standard cloud logging services.