Cloudflare and OpenAI have unveiled a powerful integration enabling developers to build intelligent, stateful AI agents that combine OpenAI’s reasoning with Cloudflare’s scalable execution infrastructure. By pairing the OpenAI Agents SDK with Cloudflare’s new Agents SDK and foundational technologies like Durable Objects and Workers, the collaboration delivers global reach, persistent memory, and human‑in‑the‑loop interaction, all within a serverless framework.
The synergy addresses a key shortcoming of stateless AI agents. OpenAI’s Agents SDK offers advanced cognition—planning, tool‑calling, decision‑making—yet leaves execution environment and persistence to the developer. Cloudflare’s solution fills that gap: its Agents SDK runs atop Workers and Durable Objects, providing each agent a unique identity, durable memory store, built‑in scheduling, WebSocket connectivity and global low‑latency execution.
Durable Objects act as the agent container. Each instantiation—based on a name or unique ID—carries its own state and storage, enabling multi‑session workflows, memory hydration, and asynchronous execution. Developers can create one agent per user, task, or domain, avoiding state entanglement while fostering modular, composable agent systems. For instance, one could build a triage agent that routes queries to specialist agents, each maintaining separate memory and logic.
A standout feature is scalability through human‑in‑the‑loop control. Cloudflare’s architecture enables agents to pause mid‑workflow, await human judgment, and resume—persisting intermediate steps and context across sessions. Knock, a third‑party messaging layer, exemplifies this. Developers have built virtual card‑issuing workflows where the AI agent pauses for approval before issuing a card—managed via Knock plus Cloudflare’s SDK.
Another innovation: agents are addressable beyond HTTP. Cloudflare’s system supports Twilio‑backed phone‑call integrations, WebSocket real‑time sessions, email and pub/sub. This opens rich, multimodal use cases—voice, text, email—bound by a globally unique agent identity.
Complementing these developments, a remote Model Context Protocol server has been introduced. Cloudflare now allows agents to host MCP servers directly, enabling structured tool integration and external service access via authenticated, remote endpoints using MCPAgent. The MCP feature dovetails neatly with Cloudflare’s recent release of a free tier for Durable Objects and general availability of multi‑step Workflows, lowering the entry barrier for developers.
Addition of the OpenAI Agents SDK and Responses API further enriches the landscape. OpenAI’s Responses API supports dynamic web search, file system access and system‑level tasks; the Agents SDK coordinates multi‑agent orchestration. Paired with Cloudflare’s persistent runtime, this empowers developers to build AI agents capable of real‑time research, memory‑backed workflows and inter‑agent communication.
Underpinning this integration is Cloudflare’s acquisition of Outerbase in April, a database platform company. The acquisition strengthens data infrastructure within Workers, Durable Objects and the Agents SDK—helping developers build rich, contextual, database‑backed AI systems. This move boosts long‑term memory storage and retrieval critical for agents maintaining evolving user context.
Industry observers are taking notice. A Medium commentary described the duo as “perfect complements: OpenAI’s Agents SDK gives you the brain, the other gives you the body”. Cloudflare’s CEO Matthew Prince emphasised that these developments remove “cost and complexity barriers” to agent deployment, calling the MCP server release “the industry’s first remote MCP server”.
Developers working with the Agents SDK can bootstrap agent projects via common workflows: installing via npm or using the agents‑starter template, extending the core Agent class to handle HTTP, WebSocket, scheduled tasks, SQL storage, and tool invocation. Integration with front‑end frameworks is supported through useAgent and useAgentChat React hooks, offering real‑time UI connections.
Looking ahead, Cloudflare promises further enhancements: evaluation tooling, voice and video interactivity via WebRTC, richer email integration for human supervision, self‑hosting capabilities, structured output support, and deeper embedding with Worker AI, Vectorize, Log Explorer and AI Gateway.
The evolving field of AI agents is entering a new phase—no longer demonstrations, but operational systems able to remember, adapt, collaborate, and operate at global scale. By combining cognitive reasoning with robust orchestration and persistence, developers are empowered to deploy production‑ready agents that are stateful, interactive and distributed.
That shift stands to redefine automation, customer support, education, workflows and more—lowering development barriers, increasing resilience, and enabling agents that truly work on behalf of users across time, platforms and modalities.