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 > <!-- deep-read:sections start --> ## 🧭 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 (state, concurrency, retries, latency, policy) are infrastructure problems. - Cloudflare supplies each missing piece — durable execution, stateful coordination, model-as-traffic gateways, edge inference, self-refreshing retrieval, governed tooling, and cryptographic agent identity. - Running the whole loop on one network removes seams and latency; in scenarios touching billing, commerce, or real-time voice, spreading across vendors becomes self-sabotage. - Cloudflare isn't competing at the commoditizing model layer; it's competing at the infrastructure layer that compounds, and has quietly built the place agents will live. ## 📰 Newsletter - For the WordPress crowd eyeing the agent wave: Remkus de Vries (prepping a Cloudflare-for-WordPress course) makes a tidy case that Cloudflare's edge stack — Workers, Workflows, Durable Objects, AI Gateway, AutoRAG, MCP — is quietly becoming the operating system for AI agents. Worth a read if you host on or build for the edge. - Learning resource angle: the piece doubles as a glossary of agent infrastructure primitives (durable execution, MCP governance, Web Bot Auth signed agents) explained in plain language — handy for site builders who keep hearing "agents" without a mental model. - WordPress tie-in: Web Bot Auth and signed-agent identity overlap directly with the agent-readiness work (Web Bot Auth + Content Signals) already landing on professional WP sites. - Helpful links: [Markdown for Agents](https://blog.cloudflare.com/markdown-for-agents/), [Durable Objects docs](https://developers.cloudflare.com/durable-objects/), [AI Gateway docs](https://developers.cloudflare.com/ai-gateway/). ## ✅ Action Items #### DevRel - Developer advocacy: build a demo agent on Workers + Workflows + Durable Objects to validate the "one fabric, fewer seams" claim with real latency numbers. - Education: workshop mapping each agent failure mode (state, retries, concurrency) to the matching Cloudflare primitive. - Developer experience: publish a quickstart template wiring AI Gateway + Workers AI + MCP for a minimal governed agent. - Developer marketing: content series "agents are distributed systems," not "smarter chatbots." - Community management: survey developers on which agent primitive they currently hand-roll (Redis, queues, cron) and why. - Internal/external/cross-project: align edge-agent guidance with WordPress agent-readiness efforts. #### OSPO - Policy & Governance: define an internal stance on single-vendor edge consolidation vs. portability before standardizing on one fabric. - License Compliance: review licensing/lock-in terms of managed primitives (AutoRAG, AI Gateway) vs. open alternatives. - Security Management: evaluate Web Bot Auth / signed-agent identity as part of a vuln + trust-boundary strategy. - Software Inventory: SBOM coverage for agent dependencies running at the edge. - Developer Support: consultation hours on durable-execution patterns and idempotency. - Community Engagement: track open standards (MCP, Web Bot Auth) and where they're vendor-neutral. - Strategic Alignment: KPI on cross-cloud hop reduction if adopting edge agents. #### Community - Maintainers: governance templates for tool/MCP access controls in shared projects. - Contributors: "first issue" path for building a sample MCP server on Workers. - Sponsors: transparent reporting on any edge-vendor dependencies introduced. - Ecosystem: coordinated disclosure drill for agent identity/trust failures. #### WordPress - Core: assess implications of agents as first-class web clients for REST API and content delivery. - Core Performance: measure token/latency impact of Markdown-for-agents-style content transformation. - Core AI: responsible-AI guardrails for agent access to WP content and actions. - **Core Program Team:** coordinate agent-readiness work across security, meta, and hosting. - Mobile: not applicable to this source — article is server/edge-infrastructure focused. - Playground: reproducible lab demonstrating an agent consuming WP content as clean Markdown. - Community: meetup talk on "what agent infrastructure means for site builders." - Training: lesson path on edge agents for the upcoming Cloudflare-for-WordPress course. - Marketing: narrative connecting WP agent-readiness to the broader agent-cloud shift. - Support: knowledge base entry on signed-agent / Web Bot Auth basics. - Design: not applicable to this source — no UX surface discussed. - Documentation: examples-first docs for exposing WP content to agents safely. - Photos: not applicable to this source. - TV: not applicable to this source. - Polyglots: not applicable to this source — single-locale infrastructure piece. - Accessibility: ensure agent-oriented content transformation preserves semantic structure used by assistive tech. - Meta: evaluate .org API behavior under agent (non-human) traffic patterns. - CLI: scaffold commands for testing WP-to-agent content delivery. - Hosting: cross-host checklist for Web Bot Auth and agent traffic governance. - Tide: not applicable to this source. - Openverse: not applicable to this source. - Themes: not applicable to this source. - Plugins: provenance/security review of plugins that expose MCP-style tool access. #### The WPCC - Programs: fellowship track on edge/agent infrastructure literacy for maintainers. - Projects: documentation sprint translating agent primitives for non-specialists. - Funding experiments: microgrant for an open, vendor-neutral agent-readiness reference. #### FAIR - Technical independence: weigh agent-cloud consolidation against local metadata/provenance and portability goals. - Policies and covenants: anti-capture clauses if standardizing on a single edge vendor. - Productized governance artifacts: scorecard rating agent infrastructure choices on lock-in vs. trust. ## 🧪 Hypotheses - If agents become first-class web clients, then sites serving clean Markdown to agents will see measurably lower agent token/latency cost than HTML-only sites. - If durable execution is bundled into one platform, then teams adopting it ship multi-step agents faster than teams assembling queues/cron/retries themselves. - If the agent loop runs on one network fabric, then end-to-end latency drops measurably versus a cross-cloud equivalent. - If model capabilities commoditize, then differentiation and margin shift toward infrastructure providers over model providers. - If cryptographic agent identity (Web Bot Auth) gains adoption, then receiving sites will increasingly gate sensitive actions on signed agents. ## 📊 Methodology - Build two functionally identical support-billing agents — one all-Cloudflare, one cross-cloud — and benchmark latency, failure recovery, and developer hours to ship. - Instrument token usage and time-to-first-action for agents consuming HTML vs. agent-Markdown versions of the same pages. - Longitudinal tracking of vendor margins and revenue mix (model layer vs. infrastructure layer) to test the commoditization claim. - Adoption telemetry for Web Bot Auth signed-agent traffic across a sample of sites over time. ### ❓ Questions - Does "coordinated" require intent, or can convergent infrastructure emerge from independent product bets? How would you tell from the outside? - Where is the lock-in cost of consolidating the whole agent loop on one vendor, and when does it outweigh the seam-reduction benefit? - Are MCP and Web Bot Auth genuinely open/portable, or do Cloudflare's implementations create soft lock-in? - Which of these primitives have credible equivalents on other clouds today? ### 🔮 Future Research - Comparative study of agent-infrastructure stacks across major clouds (AWS, GCP, Azure, Cloudflare) on the same nine capabilities. - Empirical work on how much agent latency actually comes from cross-cloud hops vs. model inference time. - Standards analysis of MCP and Web Bot Auth governance and neutrality. - Economic study of whether the infrastructure layer "compounds" relative to the model layer. ### 💡 Implications - Practice: teams should model agents as durable workflows, not request/response chatbots, and design for idempotency from day one. - Policy: organizations need a deliberate stance on single-vendor edge consolidation and agent identity before standardizing. - Security: cryptographic agent identity becomes a prerequisite, not a nice-to-have, once agents transact. - Education: "agent = distributed system + LLM brain" is a teachable frame that reorients curricula away from prompt-engineering hype. - Funding models: value (and margin) may migrate from model providers to infrastructure providers if the thesis holds. ### 🧠 Learning Styles Assessment - Reading level estimate: roughly grade 9–11 (Flesch-Kincaid Grade ~10) — short declarative sentences, some jargon (idempotency, Durable Objects, MCP) lifting it. #### 👀 Visual learners - Present: a bulleted release list and a four-line "agent loop" mapping; no diagrams or screenshots. - Improve: add an architecture diagram of the agent loop across Workers/Workflows/Durable Objects/AI Gateway, plus a before/after seams visual. #### 🔊 Aural learners - Present: none — text-only blog post. - Improve: 3-minute audio summary or TTS narration; the prose already reads aloud well. #### 🗣️ Verbal learners - Present: strong analogies ("LLM as a brain," "nervous breakdown" diagram, "models like traffic, not libraries"), clear definitions. - Improve: a short glossary box for durable execution, Durable Objects, MCP, Web Bot Auth. #### 🧩 Kinesthetic learners - Present: none — no labs or steps. - Improve: a hands-on quickstart building a minimal agent on Workers to feel the primitives. #### 🧠 Logical learners - Present: clear claim → failure-mode → primitive structure, plus three scenarios. - Improve: a decision table (failure mode → primitive → alternative you'd hand-roll) and KPIs for the latency claim. #### 🧑‍🤝‍🧑 Social learners - Present: comments section and newsletters invite discussion. - Improve: discussion prompts or a community challenge to replicate the benchmark. #### 🧘‍♀️ Solitary learners - Present: self-contained read with reflective framing ("the mental model that changes everything"). - Improve: a self-check on whether your current agent design handles retries/state. #### 🌿 Naturalistic learners - Present: organic metaphors (pipelines "rot," knowledge as a "living layer"). - Improve: lifecycle framing of context freshness as an ecosystem rather than a snapshot. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: clear H2 section headings and a logical reading order in the source. - Improve: ensure the decorative background image carries empty alt; verify heading levels aren't skipped. #### Keyboard navigation - Present: a skip link ("Skip to content") is present. - Improve: confirm comment form and newsletter links have visible focus and logical tab order. #### Contrast and readability - Present: standard blog body text, presumably meeting 4.5:1. - Improve: verify link contrast against the patterned background; check line-height for long paragraphs. #### Text resizing and reflow - Present: responsive viewport meta set. - Improve: confirm no clipping at 200% zoom and support prefers-reduced-motion (theme uses view-transitions). #### Media and captions - Present: no audio/video, so no caption gap. - Improve: if an audio version is added later, ship a transcript. #### Forms and interactions - Present: labeled comment fields (Name, Email, Website) with required markers. - Improve: inline validation and accessible error messaging on the comment form. #### Standards reference - Map gaps to WCAG 2.2 AA: 1.1.1 (non-text content), 1.4.3 (contrast), 1.4.10 (reflow), 2.4.1 (skip link, met), 2.4.7 (focus visible), 3.3.1/3.3.2 (form labels/errors). ### 🌍 DEIB Assessment #### Cultural breadth - Present: vendor-neutral technical examples (support, commerce, voice) not tied to one region. - Improve: global case studies; note that edge latency benefits vary by region/connectivity. #### Inclusive language - Present: gender-neutral, plain professional tone; jargon mostly defined in context. - Improve: expand a glossary so newcomers aren't gated by terms like idempotency or DLP. #### Neurodiversity support - Present: short paragraphs, consistent claim→evidence rhythm, frequent headings. - Improve: add a TL;DR and icon+text section markers for predictable scanning. #### Representation - Present: single-author POV; no people quoted. - Improve: cite a range of practitioners or counter-views to broaden perspective. #### Psychological safety - Present: respectful, non-attacking framing of competitors ("model layer... rapidly commoditizing"). - Improve: soften the "self-sabotage" framing, which can read as pressure rather than analysis. ### ✅ Logical Validity #### Strong reasoning - Inductive (release pattern → infrastructure intent): the eight-bullet list plus the nine-capability close build a reasonable cumulative case that the parts compose into a stack. Strong as a pattern observation; clarify by separating "these compose well" (defensible) from "Cloudflare intended this" (inference). - Analogical ("agent is a distributed system wearing an LLM as its brain"): a genuinely useful, illuminating analogy that correctly relocates the hard problems from prompting to infrastructure. Strengthen by naming where the analogy breaks (the LLM brain is non-deterministic in ways a CPU isn't). - Causal (more hops → more latency and complexity; one fabric → fewer seams): sound and well-grounded in distributed-systems reality. Clarify with actual latency figures to move from plausible to demonstrated. #### Fallacies - False dilemma — "Three scenarios where not using Cloudflare would be self-sabotage" and "distributing components across vendors becomes performance sabotage." Frames the choice as Cloudflare vs. catastrophe, ignoring competent multi-cloud and other single-vendor stacks. Fix: present it as a tradeoff with conditions where alternatives are fine. - Appeal to consequences / loaded language — "self-sabotage," "nervous breakdown," "performance sabotage" persuade via dramatized stakes rather than evidence. Fix: replace with measured comparisons and benchmarks. - Hasty generalization / mind-reading — inferring corporate intent ("starts looking coordinated") from a release cadence. Fix: hedge to "the releases compose into a coherent stack regardless of intent." - Survivorship/selection framing — the bullet list curates announcements that fit the thesis; releases that don't fit aren't weighed. Fix: acknowledge counter-evidence or non-agent releases. - Undefended assertion — "competing at the infrastructure layer, which compounds" is asserted, not shown. Fix: define "compounds" and offer evidence (margins, switching costs). #### Tone of voice - Confident, declarative, persuasive-essay register with punchy short sentences ("It was a signal."). Tech-insider but accessible; occasionally promotional given the author's Cloudflare course, though disclosed. ### 🛡️ DARVO Analysis - Deny: not applicable in the classic sense — the piece makes no accusation it must deny. The mild rhetorical move is dismissing the "scraper/nuisance" view of agents ("Not scrapers. Not nuisances.") to reposition them as legitimate clients. - Attack: no attacks on people or accusers. The closest is framing non-Cloudflare architectures as "self-sabotage" and competitors' model-layer focus as "commoditizing" — criticism of approaches, not credibility attacks. - Reverse victim/offender: not present — no accused party is recast as victim. - Emotional cues / rhetoric: stakes are dramatized ("nervous breakdown," "sabotage," "existential concerns") to motivate adoption; related fallacies are the false dilemma and appeal-to-consequences noted above. No DARVO pattern proper. ## 🎯 Conclusion - Bottom line: the piece's durable contribution is the frame — agents are distributed systems with an LLM brain, so their hard problems are infrastructure, and Cloudflare has assembled a coherent stack to meet them. The weakest move is inferring intent and overselling consolidation via "self-sabotage" language. - Next steps for readers: adopt the "durable workflow, not chatbot" mental model; benchmark the one-fabric latency and lock-in claims rather than taking them on faith; and treat agent identity (Web Bot Auth) as a real prerequisite if your agents will transact. ## 🔄 Refactor Appendix — Cloudflare is Quietly Building the Agent Cloud > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. ### 📚 Literature Atoms (H3) #### [H1] Agents are distributed systems with an LLM brain > [!quote] ID:rw_pending > An agent is not a "smart chatbot." It is a distributed system whose decision-making component happens to be an LLM. **Context & use**: - "LLM Brain" section, near the top of the argument. - Reframes agent hard problems from prompting to infrastructure; the load-bearing claim of the piece. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H2] Durable execution separates demos from systems > [!quote] ID:rw_pending > Durable execution is the line between demos and systems. **Context & use**: - "Durable execution" section. - Multi-step agent logic must survive partial failure; bundling it into the platform beats assembling queues/cron/retries. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H3] Cryptographic agent identity precedes an agent web > [!quote] ID:rw_pending > Cryptographic identity for agents is not a feature. It is a prerequisite for an agent-driven web. **Context & use**: - "Trust is the long game" section. - Once agents transact, receiving systems must distinguish authorized delegation from arbitrary automation. **Tags**: #xref #share **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H4] Infrastructure compounds while the model layer commoditizes > [!quote] ID:rw_pending > Most companies are competing at the model layer, which is rapidly commoditizing. Cloudflare is competing at the infrastructure layer, which compounds. **Context & use**: - Closing section. - The strategic claim of the essay; useful but asserted rather than evidenced. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Model an agent as a durable workflow, not a chatbot **Claim** (one sentence, Courtney's voice): An agent's reliability comes from treating it as a stateful, durable distributed system rather than a smarter prompt. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds when agents take multi-step real-world actions (billing, commerce, tool calls); less relevant for single-turn Q&A assistants. **Links**: - [[AI agents]] — because extends - [[Edge computing]] — because applies-in agent latency reduction - [[Web Bot Auth + Content Signals]] — because evidence-for agent identity as prerequisite <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Benchmark the "one fabric = lower latency" claim with a Cloudflare-only vs. cross-cloud agent — https://remkusdevries.com/cloudflare-is-quietly-building-the-agent-cloud/ #fleeting #inbox - [F2] Check whether Cloudflare's MCP / Web Bot Auth implementations are genuinely portable or soft lock-in #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Tie this article's Web Bot Auth framing into the courtneyr.dev agent-readiness notes — research #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Cloudflare Is Quietly Building the Agent Cloud.md|Cloudflare Is Quietly Building the Agent Cloud by remkusdevries.com]] ## Highlights - 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. ([View Highlight](https://read.readwise.io/read/01kvqvqq3rawe79ce8rg45tzn8)) - Note: Agents are infrastructure problems, not chatbots - Tags: atom/literature/h1 - 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. ([View Highlight](https://read.readwise.io/read/01kvqvqrhvfck84bd3h02fr275)) - Note: Durability separates real agents from prototypes - Tags: atom/literature/h2 - Cryptographic identity for agents is not a feature. It is a prerequisite for an agent-driven web. ([View Highlight](https://read.readwise.io/read/01kvqvqsyf63vdefsm3wq7qnn2)) - Note: Agent identity precedes a transacting agent web - Tags: atom/literature/h2 - 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. ([View Highlight](https://read.readwise.io/read/01kvqvqvf8ch9mxsrbhfp6x94x)) - Note: Infrastructure layer compounds vs commoditizing models - Tags: atom/literature/h2