What Is an Agentic Harness: Global Agent Infrastructure with Cloudflare Agents SDK

Buzz is social. Hermes is memory-aware. But now we must ask: what if 1 million agents run simultaneously? Each for a different customer, in a different time zone, under different network conditions?

Cloudflare Agents SDK is a harness that works at that scale.

Cloudflare is known as a CDN company. But in recent years, it has taken aggressive steps into the developer platform space—Workers (serverless compute), Durable Objects (stateful execution), D1 (SQL database), Vectorize (embedding storage), and more. All of these are the building blocks for running agents.

What Is Cloudflare Agents SDK

Agents SDK is an open-source harness written in TypeScript. It is hosted on GitHub as cloudflare/agents.

Technically: an agent runs as a Cloudflare Durable Object. Each agent has its own state store, SQLite database, and lifecycle. Like Hermes’s memory, but Hermes’s memory is bound to a single machine; Cloudflare Agents SDK’s memory is global and resilient.

Simply: “easy to start an agent, infinite to scale.”

Why Cloudflare

Cloudflare’s decision to build Agents SDK stems from business logic.

Last year, Cloudflare released Project Think, its first-party agent harness. Project Think runs hidden inside products like OpenClaw (systems automation). But Cloudflare realized thousands of customers want to build their own agents; simply open-sourcing Project Think or offering it to developers wasn’t enough, because every team uses different business rules and different tools.

Solution: open-source the harness itself, but control the platform primitives (compute, state, storage). Also ship a framework called Flue—built on Project Think, but offering a more cultivated development experience.

Result: Cloudflare Agents SDK.

Architecture: Three Layers

Remember Hermes—model binding, tools, memory.

Cloudflare structures its architecture slightly differently.

First layer: communication channels. An agent is reachable via Telegram, Discord, WhatsApp, email, web UI, webhook. It’s the same agent’s memory across all. You don’t write code to say “for this channel, do this”; the system architecture makes all channels automatically compatible.

Second layer: agent harness (agent loop). The agent calls a model, selects tools, reads tool results, decides whether to continue. Cloudflare released Project Think as an open-source harness for this. But developers can write their own harnesses; they simply use SDK primitives as the interface.

Third layer: runtime and platform. Durable Objects (stateful compute), Durable Streams (durable execution log), Workers (compute), D1 (SQL), Vectorize (embeddings), Browser API, Sandbox (code execution), and more. All unified inside the SDK and exposed to developers with a single command.

Why It Scales

In Hermes, saying “1 million agents” means 1 million Python processes. Expensive and hard to manage.

In Cloudflare, each agent is a Durable Object instance. Durable Objects enter hibernation when not in use and wake on new requests. Millions of agents can be stored as state in a database. CPU is consumed only when an agent is active, so cost is predictable and small.

Example: an e-commerce customer service system runs 100,000 customer chatbots. In Hermes, you cannot keep all 100,000 processes idle. In Cloudflare, you hibernate 100,000 Durable Objects; each chatbot wakes automatically when it receives a message.

Technical Strengths and Boundaries

Cloudflare’s strength is scale and global distribution. The moment an agent is created, it runs across 190+ countries.

Its boundary is platform lock-in. It is not self-hosted. If you want to run it on your own servers, you must reproduce Cloudflare’s architecture yourself. Even though it is open source, platform integration is Cloudflare-specific. OpenClaw (Cloudflare’s system automation product), for instance, runs only on Cloudflare.

Connection to Agentic-Organization

Buzz is social, Hermes is memory-aware, Cloudflare is scalable. But another question arises: what rules govern these 100,000 chatbots? Do they all use the same model? The same behavior set? Is there coordination between them?

This answer must be defined at an organizational level. Agentic-organization enters here. As a self-hosted, file-based platform, agentic-organization manages all agents as first-class members of an institution. While Buzz provides the social layer, Hermes supplies memory, and Cloudflare provides scale, agentic-organization creates a shared ethical framework, decision hierarchy, and control mechanism for all.

When building Cloudflare Agents SDK and asking “how do we govern 1 million agents,” the answer is found in agentic-organization. Learn more at: github.com/fab-agent/agentic-organization

Summary: Cloudflare Agents SDK is an open-source harness for global scale, with durable execution. If Hermes gains from memory, Cloudflare gains from scale. If your organization will run hundreds or millions of agents, Cloudflare is the infrastructure layer; agentic-organization is the governance layer.