Cloudflare Is Quietly Building the Agent Cloud  🔗 **Source:** [Cloudflare Is Quietly Building the Agent Cloud](https://remkusdevries.com/cloudflare-is-quietly-building-the-agent-cloud/) > [!note]+ Document Notes > --- > tags: [cloudflare, ai-agents, edge-computing, infrastructure, durable-execution, mcp, agent-identity] > author: Remkus de Vries > source: https://remkusdevries.com/cloudflare-is-quietly-building-the-agent-cloud/ > reference: https://remkusdevries.com/cloudflare-is-quietly-building-the-agent-cloud/ > date created: 2026-06-22 00:00 > --- > # 🏷️ Title: Cloudflare is Quietly Building the Agent Cloud > ## 🧭 Table of Contents > - [📝 Summary](#-summary) > - [📖 Main Content](#-main-content) > - [🗒️ Notes](#️-notes) > - [📚 References](#-references) > - [🗣️ Paraphrase](#️-paraphrase) > - [📰 Newsletter](#-newsletter) > - [✅ Action Items](#-action-items) > - [🧪 Hypotheses](#-hypotheses) > - [📊 Methodology](#-methodology) > - [❓ Questions](#-questions) > - [🔮 Future Research](#-future-research) > - [💡 Implications](#-implications) > - [🧠 Learning Styles Assessment](#-learning-styles-assessment) > - [📊 Accessibility Assessment](#-accessibility-assessment) > - [🌍 DEIB Assessment](#-deib-assessment) > - [✅ Logical Validity](#-logical-validity) > - [🛡️ DARVO Analysis](#-darvo-analysis) > - [🎯 Conclusion](#-conclusion) > ### 🔗 References > - [Markdown for Agents](https://blog.cloudflare.com/markdown-for-agents/) > - [Workflows reached general availability](https://blog.cloudflare.com/workflows-ga-production-ready-durable-execution/) > - [Building AI agents with MCP, authn, authz, and Durable Objects](https://blog.cloudflare.com/building-ai-agents-with-mcp-authn-authz-and-durable-objects/) > - [AI Gateway Aug 2025 refresh](https://blog.cloudflare.com/ai-gateway-aug-2025-refresh/) > - [AI Week 2025 wrap-up](https://blog.cloudflare.com/ai-week-2025-wrapup/) > ## 📝 Summary > - A WordPress consultant argues that Cloudflare's seemingly minor "Markdown for Agents" launch is the tell that the company is deliberately assembling end-to-end infrastructure for AI agents, not shipping unrelated features. > - It matters because it reframes the competitive frame: most vendors fight at the rapidly commoditizing model layer, while Cloudflare competes at the infrastructure layer "which compounds." > - Key stakeholders: Cloudflare, developers building agentic systems, enterprises adopting agents, model providers, and the WordPress/edge ecosystem the author teaches. > - Risks and tradeoffs: the thesis is interpretive (reading intent from a release pattern), it leans toward single-vendor consolidation, and the "self-sabotage" framing for not using Cloudflare is rhetorically loaded. > - Bottom line: agents are distributed systems whose brain is an LLM, and Cloudflare has quietly built a fast, durable, governable, trusted place for them to run. > ## 📖 Main Content > - Core thesis: an agent is "a distributed system whose decision-making component happens to be an LLM." The model proposes the next action; the surrounding system must execute it safely, reliably, repeatedly. That problem is solved by infrastructure, not better prompting. > - The author treats Cloudflare's past year of releases as a coordinated stack rather than a scatter of features: Workflows (durable execution), Durable Objects (state + identity + coordination), AI Gateway (model routing/governance), Workers AI (edge inference), AutoRAG (managed retrieval freshness), MCP servers + MCP Portals (tool governance), Firewall for AI / Secrets Store / Workers VPC (security), and Web Bot Auth (signed agents / cryptographic agent identity). > - Argument structure: each layer maps to a real distributed-systems failure mode agents hit — long-running multi-step logic that survives partial failure, concurrent tool calls corrupting shared state, RAG indexes that "rot," ungoverned tool access, and unverifiable agent actions. Keeping the agent loop (logic, orchestration, state, inference) on one fabric reduces hops, latency, and cross-cloud seams. > - Three "self-sabotage" scenarios make the case concrete: a billing-touching support agent needing durability + auditability; a commerce agent acting on a user's behalf needing cryptographic identity; a real-time voice agent needing minimal hop latency. > - Define terms: **durable execution** = multi-step logic that survives failure and resumes correctly; **Durable Objects** = stateful compute with stable identity for coordination/locking; **MCP** (Model Context Protocol) = standard for agent-to-tool access; **Web Bot Auth** = signed, cryptographically identifiable automated agents; **AutoRAG** = managed crawl/embed/store/recrawl pipeline. > - Contested/uncertain claims: the central move is inferring corporate *intent* from a release cadence ("stops looking incremental... starts looking coordinated"). That's plausible but unfalsifiable from the outside. The "competing where it compounds vs. commoditizing model layer" claim is an assertion, not evidence. The author is also a Cloudflare-for-WordPress course author, a relevant interest to note. > ### 🗒️ Notes > - Framing device: the whole piece hangs on "Markdown for Agents was not a small optimization. It was a signal." A small release reread as strategic intent. > - "An agent is a distributed system wearing an LLM as its brain" — the load-bearing mental model; everything downstream is justified by distributed-systems needs, not AI capability. > - Durability section: vivid line — without it "the architecture diagram looks like a nervous breakdown." Argues Cloudflare baked durability into the platform vs. assembling from five services. > - Durable Objects framed as "about identity and coordination" where Workflows is "about time." Contrast with frameworks that "quietly assume you will bolt on Redis." > - "Treating models like traffic, not libraries" — AI Gateway reframes inference as infrastructure/control-plane (routing, logging, spend, policy, DLP) rather than a code dependency. Pairs with Workers AI so "the agent loop shrinks." > - AutoRAG: "Most RAG pipelines quietly rot." Continuous refresh reframed as a reliability feature: "Fresh context is a reliability feature." > - MCP: "Tools are what make agents useful. Tools are also what make them risky." Governance (authn, authz, rate limiting, observability, org control) is the gap between "agent calls tools" and "enterprise allows agents to call tools." > - Trust: "Cryptographic identity for agents is not a feature. It is a prerequisite for an agent-driven web." Cloudflare "positioning itself as the trust boundary." > - Closing list reframes nine capabilities (execution, state, durability, inference, retrieval, tooling, security, identity, content transformation) as seam-elimination, not a product called "Agent Platform 1.0." > - Author context: revisiting an Obsidian note while prepping a Cloudflare-for-WordPress course. > ## 📚 References > - [Markdown for Agents — Cloudflare blog](https://blog.cloudflare.com/markdown-for-agents/) > - [Workflows GA: production-ready durable execution](https://blog.cloudflare.com/workflows-ga-production-ready-durable-execution/) > - [Building AI agents with MCP, authn/authz, and Durable Objects](https://blog.cloudflare.com/building-ai-agents-with-mcp-authn-authz-and-durable-objects/) > - [AI Gateway Aug 2025 refresh](https://blog.cloudflare.com/ai-gateway-aug-2025-refresh/) · [AI Gateway docs](https://developers.cloudflare.com/ai-gateway/) > - [Workers AI partner models](https://blog.cloudflare.com/workers-ai-partner-models/) · [Workers AI docs](https://developers.cloudflare.com/workers-ai/) > - [Introducing AutoRAG](https://blog.cloudflare.com/introducing-autorag-on-cloudflare/) · [AI docs](https://developers.cloudflare.com/ai/) > - [Remote MCP servers on Workers](https://blog.cloudflare.com/remote-model-context-protocol-servers-mcp/) > - [AI Week 2025 wrap-up (Firewall for AI, Web Bot Auth)](https://blog.cloudflare.com/ai-week-2025-wrapup/) > - [Durable Objects docs](https://developers.cloudflare.com/durable-objects/) > ## 🗣️ Paraphrase > - Markdown for Agents seems trivial — strip HTML to clean Markdown so LLMs don't burn context — but you only ship it if you expect agents to become legitimate, first-class web clients. > - Seen that way, the last year of Cloudflare releases reads as coordinated infrastructure for autonomous systems rather than a pile of separate features. > - The unifying idea: an agent is a distributed system with an LLM for a brain, so its hard problems Background pattern It was Cloudflare’s announcement of [Markdown for Agents](https://blog.cloudflare.com/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. * Durable Objects featured prominently in agent discussions around [building AI agents with MCP, auth, and Durable Objects](https://blog.cloudflare.com/building-ai-agents-with-mcp-authn-authz-and-durable-objects/). * AI Gateway expanding into routing and governance through the [AI Gateway refresh](https://blog.cloudflare.com/ai-gateway-aug-2025-refresh/) and its [developer documentation](https://developers.cloudflare.com/ai-gateway/). * Workers AI accelerating inference at the edge with [partner model expansion](https://blog.cloudflare.com/workers-ai-partner-models/) and the broader [Workers AI documentation](https://developers.cloudflare.com/workers-ai/). * AutoRAG handling ingestion and freshness as described in the [AutoRAG launch post](https://blog.cloudflare.com/introducing-autorag-on-cloudflare/) and its [AI documentation section](https://developers.cloudflare.com/ai/). * MCP servers and MCP Portals operationalizing tool access, including [remote MCP servers on Workers](https://blog.cloudflare.com/remote-model-context-protocol-servers-mcp/) and the broader AI Week coverage of MCP governance. * Firewall for AI protecting model endpoints as outlined during [AI Week 2025](https://blog.cloudflare.com/ai-week-2025-wrapup/). * Web Bot Auth introducing signed agents, also detailed in the [AI Week 2025 recap](https://blog.cloudflare.com/ai-week-2025-wrapup/). Individually, they looked like features. Collectively, they form infrastructure for autonomous systems. 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 is the line between demos and systems. When [Workflows reached general availability](https://blog.cloudflare.com/workflows-ga-production-ready-durable-execution/), 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](https://developers.cloudflare.com/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. Cloudflare’s [AI Gateway](https://developers.cloudflare.com/ai-gateway/) is one of the clearest signals of platform thinking. Instead of scattering model calls across codebases, AI Gateway centralizes them. It supports routing between providers, logging, spend tracking, policy enforcement, and DLP scanning. The refreshed version extended that control plane further. This reframes inference as infrastructure rather than as a dependency. Pair that with [Workers AI](https://developers.cloudflare.com/workers-ai/), which runs inference directly on Cloudflare’s network, and something subtle but important happens: the agent loop shrinks. Fewer hops. Fewer seams. Fewer cross-cloud integrations. Agents are chatty. They call models repeatedly between tool invocations. Every hop compounds latency and operational complexity. Keeping more of that loop on one fabric is not just convenient. It is strategically cleaner. RAG is easy to prototype and hard to maintain. Keeping indexes fresh is difficult. Handling JavaScript-heavy pages is difficult. Ensuring streaming updates reach embeddings is difficult. Most RAG pipelines quietly rot. That is why [AutoRAG](https://blog.cloudflare.com/introducing-autorag-on-cloudflare/) is more important than it looks. It collapses crawling, embedding, storage, and recrawling into a managed layer that integrates with Cloudflare’s storage and compute. When retrieval is continuously refreshed rather than periodically rebuilt, agents become less brittle. When ingestion can be paired with streaming pipelines, knowledge stops being a snapshot and starts being a living layer. Agents do not merely reason. They depend on context. Fresh context is a reliability feature. #### Tooling as infrastructure: MCP operationalized Tools are what make agents useful. Tools are also what make them risky. Cloudflare leaned heavily into the Model Context Protocol, shipping remote MCP servers on Workers, publishing multiple MCP integrations, and introducing governance through MCP Portals. The difference between “agent calls tools” and “enterprise allows agents to call tools” is governance. By operationalizing MCP rather than merely supporting it, Cloudflare moved tooling from demo feature to infrastructure layer. #### Trust is the long game Cloudflare’s DNA shows most clearly in the security layer. [Firewall for AI](https://blog.cloudflare.com/ai-week-2025-wrapup/) addresses prompt-level protection. Secrets Store centralizes credentials. Workers VPC secures private connectivity. Then there is [Web Bot Auth](https://blog.cloudflare.com/ai-week-2025-wrapup/), which introduces the concept of signed agents. If agents are going to transact, purchase, trigger workflows, and act on behalf of users, the receiving side must distinguish between arbitrary automation and authorized delegation. Cryptographic identity for agents is not a feature. It is a prerequisite for an agent-driven web. Cloudflare is positioning itself as the trust boundary. ##### 1. A support automation agent that touches billing Imagine building an agent that triages tickets, fetches account data, applies credits, and updates billing systems. It must survive retries, avoid double credits, wait for approvals, log every action, and remain auditable. Trying to stitch that across five clouds with fragile queue logic is asking for outages and accounting errors. Durable execution via Workflows, coordinated state via Durable Objects, model routing via AI Gateway, and tool governance via MCP on one fabric is simply the saner architecture. ##### 2. A commerce agent acting on behalf of users If agents begin purchasing goods or managing subscriptions, identity and trust become existential concerns. Web Bot Auth and edge-level enforcement give you a foundation for cryptographically verifiable agent actions. Building that trust layer from scratch is unrealistic. Leveraging a network that already mediates global traffic is rational. Voice agents are latency-sensitive. They call speech-to-text, reasoning models, tool APIs, and text-to-speech repeatedly within seconds. Keeping inference, orchestration, and tool execution on the same global network fabric via Workers AI, AI Gateway, and Workers dramatically reduces hop latency and integration overhead. In this context, distributing components across vendors becomes performance sabotage. Cloudflare did not launch “Agent Platform 1.0.” They extended their network with: * Execution. * State. * Durability. * Inference. * Retrieval. * Tooling. * Security. * Identity. * Content transformation. Each piece alone is unremarkable. Together, they eliminate seams. Most companies are competing at the model layer, which is rapidly commoditizing. Cloudflare is competing at the infrastructure layer, which compounds. If agents become serious actors on the internet, they will need a place to live that is fast, durable, governable, and trusted. Cloudflare has quietly built that place. Markdown for Agents was not a small optimization. It was a signal. Whether you’re here for personal reflections or deep WordPress insights, there’s more where this came from. I write two newsletters worth your time. Each serving a different part of my brain (and maybe yours too)., Cloudflare Is Quietly Building the Agent Cloud, remkusdevries.com, 01kvqvhndg3t58qmm252k8vsfp.