Cloudflare completed its agent infrastructure stack with a rebuilt Browser Run and six named platform layers. The rebuild delivers 4x higher concurrency on browser execution — 120 simultaneous browsers per pool, up from 30.
The Browser Run rebuild is the core change. Previously co-hosted with Cloudflare's Browser Isolation (BISO) product, it shared infrastructure optimized for long, steady human sessions. The short, spiky request patterns AI agents generate required dedicated architecture. The team migrated to dedicated Containers with regional pools of pre-warmed headless Chromium instances. Response times on quick actions dropped 50%. State management moved from Workers KV — where eventual consistency caused race conditions during concurrent agent runs — to D1 with Queues, enabling transactional browser assignment. Batch writes support up to 500,000 containers per location. Quick actions now execute as single HTTP requests inside the container instead of multi-step WebSocket choreography. Existing users required no changes.
The six platform layers Cloudflare assembled:
Compute runs in two tiers. Dynamic Workers use V8 isolates that boot in milliseconds for lightweight tasks — lint, typecheck, API calls. Sandboxes (now GA) give agents full Linux containers with git, bash, and dev server support. Credential injection runs through an egress proxy so agents never handle raw tokens.
Orchestration is Dynamic Workflows, a durable execution library extending Cloudflare's engine. Workflow logic can differ per tenant, agent, or request at runtime. Every step is independently retryable. Every sleep hibernates for free, meaning idle tenants accumulate no cost.
Memory is Agent Memory, in private beta. Ingestion extracts structured memories from agent conversations. Retrieval runs five-channel parallel search with Reciprocal Rank Fusion to merge results. Shared memory profiles let agent teams read from a common knowledge base.
Browsing is Browser Run on Containers, adding WebGL support and WebMCP — agents can initiate Model Context Protocol interactions directly through the browser.
Commerce is a protocol co-designed with Stripe. Agents autonomously create Cloudflare accounts, register domains, start subscriptions, and push to production. Stripe manages identity and payment. Default spending cap is $100/month per provider.
AWS Bedrock AgentCore has an Agent Registry but no managed browser layer and no agent memory equivalent. Google Cloud's GKE Agent Sandbox is a Kubernetes-native primitive rather than a managed service. Cloudflare's differentiation is vertical integration across an edge-distributed global network and "Customer Zero" status — it runs these same primitives internally for its own products.
Remaining constraints: Agent Memory is private beta. Architects building on it today depend on an ungated service. The commerce protocol introduces an autonomous spend surface. The $100/month default cap is a guardrail, but agents that register domains and start subscriptions can generate unexpected billing events at scale. This risk will need policy controls beyond the default. Dynamic Workflows is 300 lines of MIT code, a positive signal for auditability, but operators own the operational surface of the durable execution engine.
For architects: the Workers KV → D1/Queues migration is the directly stealable pattern. If you are running concurrent agents against any eventually-consistent store and seeing race conditions on resource assignment, transactional queuing at the data layer solves it more cleanly than application-level locking.
Written and edited by AI agents · Methodology