☁️ Cloudflare Agents Week — Beezus Brief
Research Date: May 19, 2026 | ~31 announcements, April 12–20 + May follow-ons | Synthesized with Opus
🖥️ Compute
🔥 Sandboxes — GA
Persistent isolated Linux micro-VMs that give any agent a real computer: shell, filesystem, background processes, installed packages, network. Start on demand, resume exactly where they left off. The @cloudflare/sandbox SDK exposes exec, gitClone, writeFile, terminal, exposePort. Full computer-in-a-Worker. Not a simulation.
🔥 Artifacts: Versioned Storage That Speaks Git
Git-compatible versioned storage built for agent scale. Create tens of millions of repos, fork from any remote Git source, hand a URL to any Git client. Every agent-generated script, config, or deployed app gets full version history with rollback. Think GitHub but purpose-built for the volume agentic code generation creates.
⚡ Durable Object Facets in Dynamic Workers
AI-generated Workers running on-the-fly can now instantiate their own Durable Objects with isolated SQLite databases. Each AI-generated app gets completely separate state — if one customer's DB breaks, nothing else is affected. Open beta on Workers Paid.
⚡ Workflows v2 — Rearchitected Control Plane
Full redesign. Now supports 50,000 concurrent workflows at 300/sec creation rate. Purpose-built for durable background agents running long-horizon tasks. Our Queues-based task system could migrate here.
⚡ Sandbox Egress Controls (Zero-Trust Outbound)
Programmable zero-trust proxy on all Sandbox outbound traffic. Inject credentials, enforce policies — agents access customer APIs without ever seeing raw keys. Critical for giving Beezus agents safe access to customer integrations (Stripe, Mailchimp, QuickBooks) without storing credentials in plaintext.
🧰 Agent Toolbox
🔥 Cloudflare Email Service — Public Beta
Agents can now natively send, receive, and process email. SPF/DKIM/DMARC auto-configure. Workers binding for direct integration, REST API with TypeScript/Python/Go SDKs, MCP server included. Inbound routing headers are HMAC-SHA256 signed to prevent spoofed callbacks. This replaces Mailgun entirely. No third-party dependency, no separate billing, no key management.
🔥 Browser Run — GA (formerly Browser Rendering)
The rename signals a real product. New: Live View (watch agents browse in real time), Human in the Loop (inject a human at captchas/decisions), Session Recordings via rrweb, Crawl endpoint returning pages as HTML/Markdown/pre-processed JSON, Quick Actions REST for screenshots/PDFs. Concurrency jumped 4x: 30 → 120 concurrent browsers per account. Native MCP for Claude Desktop/Cursor. Experimental WebMCP — sites expose structured tools agents call directly, replacing the screenshot-analyze-click loop. Rebuilt on CF Containers (May 13) for faster spin-up.
🔥 AI Search — Open Beta (formerly AutoRAG)
Hybrid retrieval built for agents: semantic (vector embeddings) + keyword (BM25 with Porter/trigram tokenization) + relevance boosting. ai_search_namespaces binding, storage in R2, vectors in Vectorize, crawling via Browser Run. Free during beta. Replaces our manual Vectorize + D1 RAG architecture with something better out of the box.
⚡ Agent Memory — Managed Service
Persistent memory across sessions as a managed service. Agents recall relevant context, auto-forget stale info, get smarter over time. No manual vector DB or D1 schema. Plugs directly into the Agents SDK. Replaces the memory layer we currently maintain ourselves.
⚡ Voice SDK — Experimental
Real-time voice over WebSockets. STT via Deepgram Flux + Nova 3, TTS via Deepgram Aura. withVoice(Agent) wraps any existing agent in ~30 lines of server code. React hooks on the client. Beez can talk. The entire onboarding interview could become a voice conversation.
⚡ Unified Inference Layer (AI Gateway v2)
AI Gateway now covers 70+ models across 12+ providers: Anthropic, OpenAI, Google, Alibaba, Bytedance, RunwayML, AssemblyAI, and more. Workers AI binding works for all third-party models. One endpoint for everything. Multimodal catalog significantly expanded.
💡 Project Think — Agents SDK v2 Preview
Preview of next major SDK version. Moving from lightweight primitives to batteries-included — native reasoning loops, persistence, and memory built in. Not GA yet but signals clearly where CF is heading.
💡 Unweight — 22% LLM Compression
Lossless inference-time compression, 22% smaller model footprint, no quality loss. Cheaper and faster Workers AI inference. No action required — we benefit automatically at the infrastructure level.
🚀 Agents Creating Real Infrastructure
🔥 Agents Can Create CF Accounts, Buy Domains, Deploy — April 30
The jaw-dropper. Agents can now programmatically search domain availability, purchase domains via the Registrar API, provision new CF accounts, and deploy Workers — all from inside another Worker. A Stripe partnership handles identity and payment tokenization with a default $100/month spending cap per agent. An agent can go from "customer needs a website" to a live deployed site with a real domain with zero human involvement.
⚡ Registrar API
Domain search, availability check, and registration directly from a Worker or the cf CLI. Powers the above. Beezus agents can provision client-facing domains as part of onboarding or Wish fulfillment.
⚡ cf CLI — Unified
New unified CLI replacing all fragmented per-product CLIs. Consistent interface across ~3,000 CF API operations. Local Explorer for debugging KV/D1/R2. Both humans and agents can use it programmatically.
🔒 Security
⚡ Cloudflare Mesh
Unified private networking connecting users, servers, agents, and Workers into one secure fabric. Workers VPC integration — agents access private databases and internal APIs without manual tunnels. 5-minute setup. Think Tailscale but native to CF and agent-aware.
⚡ Managed OAuth for Access (RFC 9728)
Agents authenticate to internal apps on behalf of human users via standard OAuth delegation. No more service accounts. Agents get proper delegated identity. One click to make any internal app "agent-ready." Critical for giving Beezus agents safe access to customer tools without storing raw credentials.
⚡ Non-Human Identity Security
Scannable API tokens, enhanced OAuth visibility, GA of resource-scoped permissions. Full least-privilege architecture for multi-agent credential management. Prevents credential leakage across the hive.
💡 MCP Reference Architecture for Enterprise
CF's internal playbook for governing MCP at scale: Access + AI Gateway + MCP server portals. "Code Mode" for token cost reduction. Shadow MCP detection in Gateway. Directly relevant since Charlie already has an MCP worker deployed.
🛠️ Developer Experience
💡 Agent Lee (Beta)
AI assistant embedded in the CF dashboard. Uses sandboxed TypeScript to execute actions. Natural language for debugging Workers, inspecting KV, diagnosing routes. Our CF infrastructure co-pilot — usable right now while building.
💡 Flagship — Native Feature Flags
Feature flag service built on KV + Durable Objects. Sub-millisecond evaluation with no third-party round-trips. Replaces our ad-hoc flag system with per-request decisions at the edge.
💡 Claude Managed Agents on Cloudflare — May 19
Announced today. Anthropic's Claude running as managed agents on CF infrastructure with enhanced security, observability, and CF's networking layer baked in. Direct Workers integration.
🔲 PlanetScale + Workers
Postgres/MySQL on CF Workers. Not relevant — we're on D1 and it's working fine.
🎯 Top 5 Highest-Leverage Integrations for Beezus
Specific to our architecture. Ranked by impact × feasibility.
1. 🔥 Drop Mailgun. Ship CF Email. 1–2 weeks
We literally ran into this today. CF Email auto-configures SPF/DKIM/DMARC, has a Workers binding, and includes an MCP server. Every agent — Charlie sending sequences, Beez sending updates, Winston sending wish notifications — gets native email with zero third-party dependency. The HMAC-signed inbound routing also unlocks email-reply-as-action: customer replies to Charlie's outreach, it automatically updates the CRM stage. No Mailgun billing, no key management.
2. 🔥 Sandboxes → Winston's Wish Store Goes From Catalog to Live Deployments 4–6 weeks
Right now the Wish Store is a catalog of automations customers can activate. With Sandboxes GA, Winston can actually build and deploy them. Customer says "I want automated follow-up emails after every showing" → Winston spins a Sandbox, writes the Worker, tests it, deploys it, stores it in Artifacts with version history, hands the customer a live running automation with a rollback button. The leap from "AI that describes work" to "AI that ships work."
3. 🔥 Agents Creating Accounts + Registrar API → Hana Builds Real Client Infrastructure 4–8 weeks
When a customer needs a web presence, Beezus provisions it. Landing page, domain, client portal — Hana hires the right specialist, they use the Registrar API + Workers deploy to deliver a live URL, autonomously. Also massive for Enrollment Mentor: every college that signs up gets their own subdomain provisioned programmatically with their custom RAG loaded from uploaded materials. No DevOps required on their end.
4. 🔥 Browser Run Crawl Endpoint → Onboarding Enrichment Gets Real 1 week
Right now runOnboardingEnrichment does basic industry detection. With the Crawl endpoint returning clean Markdown/JSON, we scrape the customer's actual website during onboarding — products, pricing, team bios, testimonials, brand signals — and feed ALL of it to Charlie, Ivy, Hana, Frankie as structured context. The hive builds around the real business. Human-in-the-Loop means if the site has a login wall or CAPTCHA, it pauses and asks the customer instead of silently failing.
5. 🔥 AI Search → Replace Manual Vectorize + D1 RAG Architecture 2–3 weeks
We currently maintain our own vector storage, chunking, and retrieval logic across D1 and Vectorize. AI Search handles all of that with hybrid semantic + keyword retrieval, relevance boosting, and R2 storage — free during beta. Replace the entire RAG layer for agent memory, customer knowledge bases, onboarding transcripts, brand DNA, documents. Every agent query gets better retrieval than anything we've hand-tuned.
⚡ Quick Wins — Ship This Week
| # | What | Effort | Payoff |
|---|
| 1 | Replace Mailgun with CF Email | Medium | Eliminate vendor dependency, save money, unlock inbound email workflows |
| 2 | Browser Run Crawl in onboarding enrichment | Small | Scrape customer website during onboarding — far richer context from day one |
| 3 | Upgrade to Workflows v2 | Small | 50K concurrent durable tasks — migrate our Queues-based task system |
| 4 | Flagship for feature flags | Small | Replace ad-hoc flag system with sub-millisecond CF-native edge flags |
| 5 | Agent Lee for our dev workflow | Zero | Just start using it — debug beezus-worker via natural language in the CF dashboard |
🚀 Bigger Swings — Roadmap Movers
"Deploy My Wish" — Winston + Sandboxes + Artifacts
Winston goes from wish curator to wish builder. Customer activates a Wish → Winston spins a Sandbox, writes the automation Worker, tests it, stores it in Artifacts with full version history, deploys it live. Customer gets a running automation with a rollback button. The Wish Store becomes a delivery engine, not a catalog. This is the product leap from "AI that recommends" to "AI that ships."
Beezus as Infrastructure Layer — Agents Provisioning Real Businesses
When a customer needs a web presence, Beezus builds and deploys it. Landing page, booking portal, client-facing subdomain — Hana hires the right specialist, the specialist uses the Registrar API + CF deploy to deliver a live URL. For Enrollment Mentor: every college that signs up gets a provisioned subdomain with their custom RAG loaded from their materials. Beezus stops being a tool businesses use and starts being the infrastructure their business runs on.
Voice-First Beez
30 lines of server code. Beez can talk. The onboarding interview — the thing we've been fighting to get right all week — becomes a voice conversation on mobile. A nervous business owner talking to their new AI business partner instead of typing into a chat box. For Enrollment Mentor, the "mentor in your pocket" becomes completely literal.
CF Email as Charlie's Full CRM Backbone
Charlie owns sequences. With CF Email's inbound HMAC routing, Charlie sends from the customer's branded address, receives replies, detects intent (interested/objection/unsubscribe), and automatically updates deal stages. Full email-native CRM — no Mailgun, no Zapier, nothing outside the CF ecosystem. The sequence → reply → stage-update loop becomes a single Worker.
AI Search → Every Agent Has a Real Brain
Replace the manual Vectorize setup across the board. Agent memory, customer knowledge bases, onboarding transcripts, brand DNA, uploaded documents — all under AI Search's hybrid retrieval. Hana finds the right specialist faster. Charlie surfaces the right contacts. Beez remembers what a customer said six months ago. Every agent gets smarter without us maintaining any RAG infrastructure.
The one-line summary: Cloudflare just made it possible for Beezus agents to send email, browse the web, run code, deploy apps, buy domains, and remember everything — natively, on the same infrastructure we're already running. The platform didn't just add features. It removed the ceiling.