Cloudflare is quietly building the agent cloud - Remkus de Vries

Skip to content

Remkus de Vries on YouTube

Remkus de Vries' profile on Instagram

Cloudflare is quietly building the agent cloud

22nd of June, 2026

It was Cloudflare’s announcement of Markdown for Agents a few months bac that made everything click for me. I’m revisting the note I had saved in my Obsidian vault in preparation for my upcoming Cloudflare for WordPress course and decided to write out what I learned so far on Cloudflare’s intentions as I see them.

On the surface, the Markdown for Agents announcement looks like a small optimization. Convert messy HTML into clean, token-efficient Markdown so LLMs don’t waste context on navigation menus, JavaScript, and layout debris. That’s sensible. Almost boring.

But you don’t ship a feature like that unless you believe agents are about to become first-class consumers of the web. Not scrapers. Not nuisances. Not traffic anomalies to block. Legitimate clients.

And once that thought settles in, the past year of Cloudflare announcements stops looking incremental. It starts looking coordinated.

Workflows going GA via the announcement that Workflows reached general availability.

Durable Objects featured prominently in agent discussions around building AI agents with MCP, auth, and Durable Objects.

AI Gateway expanding into routing and governance through the AI Gateway refresh and its developer documentation.

Workers AI accelerating inference at the edge with partner model expansion and the broader Workers AI documentation.

AutoRAG handling ingestion and freshness as described in the AutoRAG launch post and its AI documentation section.

MCP servers and MCP Portals operationalizing tool access, including remote MCP servers on Workers and the broader AI Week coverage of MCP governance.

Firewall for AI protecting model endpoints as outlined during AI Week 2025.

Web Bot Auth introducing signed agents, also detailed in the AI Week 2025 recap.

Individually, they looked like features.

Collectively, they form infrastructure for autonomous systems.

LLM Brain

An agent is a distributed system wearing an LLM as its brainThis is the mental model that changes everything.

An agent is not a “smart chatbot.” It is a distributed system whose decision-making component happens to be an LLM. The model proposes the next action. The rest of the system must execute it safely, reliably, and repeatedly.

That means maintaining real state across sessions. It means coordinating concurrent actions. It means retrying external calls without duplicating side effects. It means waiting for webhooks and resuming correctly. It means observing behavior in production. It means enforcing policy at the network edge. It means handling latency across regions.

None of that is solved by better prompting. It is solved by infrastructure.

Cloudflare, perhaps more than anyone else right now, has been building that infrastructure in a way that assumes agents are not experiments but actors.

Durable execution

Durable execution is the line between demos and systems. When Workflows reached general availability, it did not generate hype. It should have.

Durable execution means multi-step logic survives failure. If an agent analyzes a ticket, calls a CRM, waits for approval, updates billing, and then notifies the customer, that is not a single request. It is a sequence that can stretch over minutes or hours and will inevitably encounter partial failure.

Without durability, you end up stitching together queues, cron jobs, retries, and idempotency checks until the architecture diagram looks like a nervous breakdown.

With durability, you model long-running logic as a first-class construct.

That is the difference between an assistant that answers questions and an agent that performs work.

Cloudflare built durable execution directly into the platform instead of expecting developers to assemble it from five services.

You can read the announcement as a product launch. Or you can read it as a signal that Cloudflare understands agents are workflows, not prompts.

State that actually persists: Durable Objects as an agent primitive

If Workflows is about time, Durable Objects are about identity and coordination.

Durable Objects provide stateful compute with a stable identity. In agent terms, that gives you a place where a user’s session can live. A place where an agent instance can coordinate actions. A place where locks can prevent concurrent tool calls from corrupting shared state. A place where you can keep a ledger of what has already been done.

Most agent frameworks quietly assume you will bolt on Redis and hand-roll coordination logic. Cloudflare gives you a substrate for it.

When agents begin to orchestrate real systems, that substrate becomes indispensable.

Treating models like traffic, not libraries

Cloudflare’s AI Gateway is one of the clearest signals of platform thinking. Instead of scattering model calls across...