Cloudflare Completes Its Agent Infrastructure Stack with Browser Run Rebuild and Six Layer Platform

Justy Okay so Cloudflare rebuilt Browser Run from scratch and I'm actually kind of impressed by the specifics here.

Cody Impressed by the specifics or impressed that they finally fixed a capacity problem they caused themselves?

Justy Both? The numbers are wild. One hundred twenty simultaneous browsers, four times what they had before. Fifty percent faster response times. And they did it by basically untangling two products that were fighting each other for resources.

Cody Right. So Browser Run was sharing infrastructure with their Browser Isolation product — BISO. And human sessions versus AI agent requests are fundamentally different workloads. Long, steady sessions versus short, spiky bursts.

Justy Exactly. And the article says that spike forced the rethink. They couldn't keep pretending those were the same thing.

Cody Mm-hm.

Cody The technical fix is interesting though. They moved state management from Workers KV — which has eventual consistency — over to D1 with Queues. And the reason that matters is race conditions. If two agent requests both think they own the same browser session, things break fast.

Justy So D1 with Queues gives them transactional assignment. The article says batch writes handling up to five hundred thousand containers per location. That's not a small number, Cody.

Cody It's not. Though I want to know what "per location" actually means in practice. Cloudflare has a lot of locations but it's not a global guarantee. But okay, the architecture makes sense. They also killed the WebSocket choreography for quick actions and replaced it with a single HTTP request that runs entirely inside the container.

Justy That's the part that gets you the fifty percent faster response times. Less round-trip overhead, no WebSocket handshake, just a direct HTTP call into the container.

Cody Sure, but here's where I get skeptical. This is a rebuild of one layer. And Cloudflare is framing it as the completion of a full agent infrastructure stack. Six layers, they say.

Justy Yeah, and I actually want to hear you walk through those because some of them sound genuinely useful. Compute, orchestration, memory, browsing, commerce.

Cody Okay so compute comes in two tiers. Dynamic Workers — V8 isolates, millisecond boot times, good for lightweight stuff like linting and API calls. And Sandboxes — full Linux containers, generally available now, with secure credential injection through an egress proxy so agents never see raw tokens. That part is actually thoughtful.

Justy The egress proxy thing is a real security win. If you're building agents that touch cloud resources, you don't want them holding live credentials.

Cody Agreed. Then orchestration is Dynamic Workflows. Roughly three hundred lines, MIT licensed, extends their durable execution engine. Every step independently retryable, sleeps hibernate for free, idle tenants cost basically nothing.

Justy That's a clean developer experience pitch. Three hundred lines is the kind of thing that makes a developer think "I could read this."

Cody Exactly. Now memory is Agent Memory and it's in private beta. Dual-pass ingestion pipeline for extracting structured memories from conversations, five-channel parallel search with something called Reciprocal Rank Fusion.

Justy Reciprocal Rank Fusion — that's a retrieval ranking technique, right? You're combining multiple search channels and re-ranking based on reciprocal rank scores.

Cody You've been reading. Yeah, it's a known approach. The five-channel parallel bit is interesting — you're running multiple retrieval strategies simultaneously and then fusing the results. It's a solid design for a managed memory service.

Justy It's in private beta though. So take it with a grain of salt until people actually use it.

Cody Oh definitely. And then commerce — they co-designed a protocol with Stripe. Agents can autonomously create accounts, register domains, start subscriptions. Stripe handles identity and payment, with a one hundred dollar per month default spending cap per provider.

Justy That spending cap is the right call. You do not want an agent running up a bill because of a loop or a bad prompt.

Cody Agreed. Now here's where I push back on the article's framing. It asks whether this is "the most complete agent infrastructure offering outside the hyperscalers." And the answer is... maybe? But completeness across layers isn't the same as depth in any single layer.

Justy That's fair. AWS has Bedrock with AgentCore and Agent Registry but no managed browser and no agent memory equivalent. Google Cloud has GKE Agent Sandbox but it's Kubernetes-native, not a managed platform. So Cloudflare does cover more layers out of the box.

Cody Right, but AWS and GCP have years more maturity in the pieces they do have. And "most complete" for agent workloads is a narrow window right now. Ask me again in a year when people have actually stress-tested Agent Memory and Dynamic Workflows in production.

Justy That's a fair skept position. I just think the Stripe integration and the browser layer together are a combo that nobody else is offering as a managed service. That might matter more than the depth argument.

Cody It might. I'm not saying it's nothing. I'm saying the marketing is ahead of the evidence.

Justy There it is. The Cody special. "The marketing is ahead of the evidence."

Cody I can't help it, Justy. You know how I work.

Justy Anyway — if you're building agent workflows today and you want the whole stack managed by one vendor, this is worth a look. Especially if you're already on Cloudflare for other things. The browser layer alone solves a problem that otherwise requires you to roll your own infrastructure.

Cody Agreed. And watch the Sandboxes GA — the egress proxy credential injection is the feature I'd actually want to dogfood first.

Justy Good call. Alright, that's Episode four twenty-eight. Cody, stop being so skeptical for five minutes.

Cody I'll try. No promises.