Cloudflare and OpenAI Release New Agent Runtimes and SDKs, Shifting the Web’s AI Infrastructure

On April 15, Cloudflare launched Project Think, a new Agents SDK that introduces durable execution, crash recovery, checkpointing, isolated sub‑agents, persistent tree‑structured sessions, and sandboxed code on Dynamic Workers. The same day, OpenAI released its next‑generation Agents SDK featuring native sandbox execution and a model‑native harness, positioning both companies to answer the question of how long‑running AI agents can be reliably run in production. Cloudflare followed up on April 16 with five additional components: AI Platform, a vendor‑agnostic inference layer routing models for agents; AI Search, a managed vector index and chunking pipeline for agent‑side RAG; Email Service beta, PlanetScale Postgres and MySQL inside Workers; and infrastructure for hosting large open‑source LLMs like Kimi K2.5 directly on its network. The shift moves the focus from the model itself to the runtime that fetches, parses, executes JavaScript, negotiates authentication, and supplies structured data to the model. Web developers must now ensure their endpoints return machine‑readable structured responses, support multi‑call session authentication, and maintain structured data integrity when JavaScript is not executed. The new runtime architecture will become the default interface for all future AI agents, making runtime‑legibility a prerequisite for AI search and commerce. Agencies and enterprises should audit existing sites for runtime compatibility and adjust authentication scopes accordingly. The rollout signals a broader industry move toward durable, auditable, and sandboxed agent execution on the edge.