From Shared Infrastructure to Container-Native Browser Run
Cloudflare has rebuilt Browser Run on its own containers platform, transforming it from a shared service into a container-native layer tuned for AI workloads. Previously, Browser Run shared infrastructure with Browser Isolation, a product optimized for long-lived human browsing sessions. That model struggled under the spiky, short-lived traffic patterns of AI agents, which pushed request volumes beyond existing capacity. By moving to dedicated Cloudflare containers, the team unlocked 4x higher concurrency, scaling from 30 to 120 simultaneous browser instances per location. This redesign also introduced regional pools of pre-warmed browsers, reducing cold-start penalties and smoothing out bursty traffic. For developers, this means Browser Run performance is now aligned with how AI agents actually behave: lots of small, fast interactions instead of single, long sessions. The result is a browsing substrate that can keep pace with modern agent orchestration and parallel task execution.
Architectural Changes Driving 50% Faster Response Times
The concurrency gains in Browser Run are matched by a 50% improvement in response times for quick actions, driven by several deep architectural changes. State management migrated from Workers KV, which is eventually consistent and prone to race conditions under heavy parallelism, to a combination of D1 and Queues. D1 provides transactional assignment of containers, while Queues batch writes to coordinate up to 500,000 containers per location. On the execution path, the team removed the previous multi-step WebSocket choreography required to perform simple browser actions. Instead, quick interactions are now handled as single HTTP requests fully executed inside the container, minimizing protocol overhead and round trips. For developers building AI agents that issue many short browser tasks—such as scraping, checking DOM state, or rendering snippets—this concurrency optimization translates directly into lower latency, higher throughput, and more predictable Browser Run performance under load.
Completing a Six-Layer Agent Infrastructure Stack
The Browser Run rebuild is not an isolated upgrade; it completes what Cloudflare positions as a six-layer platform for AI agent infrastructure: compute, persistence, networking, orchestration, observability, and security. On the compute side, Dynamic Workers offer ultra-fast V8 isolate execution for lightweight tasks, while Sandboxes provide full Linux containers when agents need tools like git, bash, and development servers, with secure credential injection via egress proxy. For persistence and memory, Agent Memory uses a dual-pass ingestion pipeline and five-channel parallel search with Reciprocal Rank Fusion to give agents structured, retrievable context, including shared memory profiles for teams of agents. Orchestration relies on Dynamic Workflows, an MIT-licensed library that lets workflow logic vary per tenant or request while leveraging durable execution, fine-grained retries, and cost-efficient hibernation. Browser Run becomes the browsing layer, now enhanced with WebGL and WebMCP support to integrate Model Context Protocol workflows directly from the browser.
Containers, Commerce, and a Vertically Integrated Agent Platform
By making Browser Run container-native on Cloudflare containers, the platform now offers a more resource-efficient and scalable substrate for AI agent workloads. Regional pools of pre-warmed containers, transactional state coordination, and single-request execution paths combine into an environment where compute, networking, and storage are tightly integrated across the edge. Beyond browsing, Cloudflare also adds a commerce layer via a protocol co-designed with Stripe, enabling agents to autonomously create accounts, register domains, start subscriptions, and deploy to production. Stripe supplies identity and payment handling, while Cloudflare focuses on infrastructure and deployment flows. Together, these elements position Cloudflare as a comprehensive platform for agent deployment and management, distinct from hyperscaler offerings like AWS Bedrock AgentCore or Google’s Kubernetes-based agent primitives. For engineering teams, this stack offers a unified surface where AI agents can compute, remember, browse, orchestrate, and transact, all on a single, globally distributed platform.
