Cloudflare has rebuilt Browser Run on top of its own Containers platform, delivering 4x higher concurrency (120 simultaneous browsers, up from 30), 50% faster response times for quick actions, and support for WebGL and WebMCP.
The upgrade, which required no changes from existing users, caps a sequence of launches over the past two months that collectively assemble a full-stack platform for building and running AI agents. Ruskin Constant, Rui Figueira, and Sofia Cardita from the Browser Run team write:
AI agent builders discovered Browser Run and quickly brought request volumes outpacing our existing capacity.
That demand spike forced an architectural rethink. Browser Run previously shared infrastructure with Cloudflare's Browser Isolation (BISO) product, where long, steady human sessions clashed with the short, spiky usage patterns of AI agents. The rebuild moved to dedicated Containers with regional pools of pre-warmed browsers, migrated state management from Workers KV (eventual consistency causing race conditions) to D1 with Queues (transactional assignment, batch writes handling up to 500,000 containers per location), and eliminated the multi-step WebSocket choreography for quick actions in favor of single HTTP requests executed entirely inside the container.
Browser Run is the browsing layer, but the broader story is the platform. Over the past two months, Cloudflare has shipped six distinct infrastructure primitives that together cover every layer an agent platform needs:
- Compute comes in two tiers. Dynamic Workers provide V8 isolate-based execution that boots in milliseconds for lightweight tasks such as lint, typecheck, and API calls. Sandboxes, now generally available, provide full Linux containers for when agents need git, bash, dev servers, and multi-language builds, with secure credential injection via egress proxy so agents never see raw tokens.
- Orchestration is handled by Dynamic Workflows, a roughly 300-line MIT-licensed library that extends Cloudflare's durable execution engine so workflow code can differ per tenant, agent, or request at runtime. Every step is independently retryable, every sleep hibernates for free, and idle tenants cost approximately nothing.
- Memory is provided by Agent Memory, currently in private beta. The service extracts structured memories from agent conversations using a dual-pass ingestion pipeline and retrieves them through five-channel parallel search with Reciprocal Rank Fusion. Shared memory profiles let teams of agents access common knowledge.
- Browsing is now Browser Run on Containers, giving agents headless Chromium instances they can control via the DevTools Protocol or the Agents SDK. The WebMCP addition enables Model Context Protocol interactions directly through the browser.
- Commerce rounds out the stack. Through a protocol co-designed with Stripe, agents can autonomously create Cloudflare accounts, register domains, start subscriptions, and deploy to production. Stripe handles identity and payment with a $100/month default spending cap per provider.
The competitive question is whether this constitutes the most complete agent infrastructure offering outside the hyperscalers. AWS has Bedrock AgentCore with Agent Registry, but no managed browser or agent memory equivalent. Google Cloud has GKE Agent Sandbox, but as a Kubernetes-native primitive rather than a managed platform service. Neither offers anything comparable to the commerce protocol.
Cloudflare's bet is different from both: a vertically integrated stack where every layer is edge-distributed across its global network, every binding is heading toward a dynamic per-tenant counterpart, and the platform itself is what Cloudflare calls "Customer Zero" for, running its own products on the same infrastructure it sells. Whether this integration advantage outweighs the hyperscalers' broader ecosystem remains an open question for platform teams evaluating where to build their agent infrastructure.
Browser Run improvements are live now on all Workers plans. The Agents SDK includes built-in Browser Run support.