Rivet: Orchestrate AI Agents Using S3 as Your Data Plane

Rivet, a Y Combinator-backed open-source framework, launched this week to scale and orchestrate AI agents using a novel approach: treating Amazon S3 as the primary data plane. The project surfaced on Hacker News with a launch thread discussing its architecture for cost-effective, durable agent workflows.

Tool: Rivet | Core Tech: S3 State Backend | License: MIT
Language: TypeScript/Node.js | Available: GitHub, Cloud

How S3 Becomes an Agent Orchestrator

Traditional agent frameworks like LangChain or LangGraph manage workflow state in memory or a dedicated database, which scales expensively. Rivet’s core innovation is persisting every agent state transition directly to S3. Each step in a chain or graph is a discrete object stored in the cloud storage service.

This turns S3 into a durable, auditable log. The framework handles the complexity of reading, writing, and coordinating these states, allowing developers to define agent graphs in code while Rivet manages execution and persistence.

The Cost and Scale Proposition

The primary benchmark is cost. Rivet’s documentation claims this architecture can reduce orchestration costs by over 90% compared to running stateful orchestration engines in memory on services like AWS Lambda or in containers.

The system is designed for long-running, asynchronous agents. A research agent that runs for hours, periodically waking up to process new data, incurs minimal cost when idle, as state is cheaply stored in S3 versus an always-on container.

How to Try It: Local and Cloud

Implementation is through a TypeScript SDK. The quickest start is local simulation.

git clone https://github.com/rivet-dev/rivet
cd rivet/examples
npm install
npm run simulate

This runs a predefined agent graph locally, but still persists state to a local file simulating S3. For cloud deployment, you configure an S3 bucket and AWS credentials. The Rivet runtime service can be deployed as a Docker container on ECS, Kubernetes, or even as a serverless function that processes steps from an S3 event queue.

import { Graph, State } from '@rivet-gg/rivet';
const graph = new Graph();
const llmStep = graph.addStep('llm', { model: 'gpt-4' });
const toolStep = graph.addStep('calculator', { operation: 'add' });
graph.addEdge(llmStep, toolStep);
// Execution is handled by Rivet's runtime
const result = await rivet.run(graph, { input: "What is 123 + 456?" });

Pros, Cons, and Early Feedback

Advantages:

• Radical Cost Reduction: For high-volume or long-duration agents, moving state out of compute is the primary financial benefit.
• Built-in Audit Trail: Every execution is automatically logged to S3 for replay, debugging, and compliance.
• Serverless Native: Fits naturally into event-driven AWS architectures (Lambda, EventBridge).

Limitations and Concerns:

• Latency Overhead: Each step incurs S3 read/write latency (~100-200ms). This makes it unsuitable for real-time synchronous agents requiring sub-second responses.
• Vendor Lock-in: Deeply tied to AWS S3. While the state abstraction could be adapted, it's currently a single-cloud design.
• Early Stage: As a new open-source project, the ecosystem of integrations and community support is limited compared to established frameworks.

Hacker News commenters noted the cleverness of the approach but questioned complexity: "This is essentially building a distributed state machine on top of S3. Powerful, but now you have to debug your agents via S3 logs," one user observed.

Alternatives and When to Choose Rivet

The agent orchestration space has clear leaders. Your choice depends on workflow characteristics.

Choose Rivet if: You are building asynchronous AI agents (e.g., research assistants, content generation pipelines, batch processing bots) where cost and durability are primary concerns, and latency is not critical.

Avoid Rivet if: You need low-latency, synchronous agent interactions (e.g., customer support chatbots, real-time gaming NPCs) or are not committed to the AWS ecosystem.

Bottom line: Rivet trades execution latency for dramatic cost savings and durability, making it a niche but powerful tool for specific production agent workloads.

Who Should Use This Now?

Rivet is immediately practical for startups and teams already on AWS deploying non-real-time AI agents at scale. The cost savings are tangible for workflows that run for minutes or hours. It's also a compelling option for ML engineers tasked with operationalizing research prototypes into durable, auditable pipelines.

Individual developers prototyping simple agents may find the S3 overhead unnecessary; LangGraph or CrewAI offer faster iteration for local, in-memory graphs. Rivet's value unlocks at the point where agent compute costs become a line item.

The framework represents a maturation in agent infrastructure, treating orchestration as a data persistence problem rather than just a compute one. As agents move from demos to production, tools that prioritize cost and auditability alongside capability will define the next stack.

The shift towards externalized, cheap state suggests a future where the most expensive part of an agent—its memory and reasoning trace—becomes its cheapest and most durable asset. Rivet is an early bet on that architecture.