Brad's Google Gemini Research · July 2026
Corporate history, V8 Isolates, storage primitives, the autonomous AI agent stack (Think, Browser Run, x402), enterprise blueprints, and compliance guardrails.
These are materially new capabilities, details, or concepts not present in the Cloudflare Playbook research.
An opinionated, stateful agent harness built on top of the Agents SDK. The playbook covered the Agents SDK broadly, but Think is a specific higher-level framework that auto-wires:
Why this matters for us
If we build AI agents on Cloudflare, Think is the framework to use. It's a more opinionated, battle-tested layer than raw AIChatAgent. The capability-based authorization model (Tiers 0-4) is exactly the security approach we need for client-facing agents.
The playbook mentioned "Browser Rendering" briefly. Gemini uncovered significantly more detail on the agent-facing Browser Run system:
Critical for our Puppeteer migration
The Human-in-the-Loop feature is exactly what we need for Facebook login. Instead of Brad launching a visible Chrome window on the Mac Mini, an agent on Cloudflare could pause at the login screen, send Brad a link, he logs in on his phone, and the agent resumes. This could eliminate our entire Mac Mini browser dependency for social media posting.
A structured permission model for what agents can do, escalating from safe to powerful:
| Tier | Capability | What It Enables |
|---|---|---|
| Tier 0 | Workspace Directory | Local file operations (create, edit, search) backed by SQLite + R2 |
| Tier 1 | Dynamic Worker | Write and run custom code in a sandboxed V8 Isolate (zero network access) |
| Tier 2 | Runtime npm Packages | Dynamically install npm libraries via @cloudflare/worker-bundler, compiled on-the-fly |
| Tier 3 | Web Automation | Spin up headless Chrome browser to interact with websites lacking APIs |
| Tier 4 | Sandbox | Full secure OS container: clone repos, compile code, run test suites |
Architectural significance
This is how Cloudflare thinks about agent security. Each tier is explicitly granted. An agent starts at Tier 0 and only gets browser access (Tier 3) or full container access (Tier 4) if you explicitly enable it. This maps well to building client-facing agents where you want tight control over what they can do.
A deferred payment protocol built with Coinbase that lets autonomous agents purchase resources (API access, datasets, compute, browser rendering hours) on the fly using micropayments. Processed via Stripe or stablecoin networks. Website owners can block scraper bots or allow them in exchange for automated x402 micropayments.
Forward-looking
Not immediately actionable for us, but this is the infrastructure for agents that can autonomously pay for services. Relevant to the "AI agent products for staffing clients" vision where agents might need to purchase data, API calls, or compute.
An open standard that lets websites explicitly declare their available APIs and actions for AI agents. Instead of agents scraping visual layouts (brittle), websites publish a machine-readable manifest of what agents can do. Agents discover and interact with sites directly through structured endpoints.
Relevant to our MCP architecture
We're already running MCP servers locally on the Mac Mini. WebMCP is the Cloudflare-hosted version of this concept for the broader web. As adoption grows, our agents could interact with third-party services through WebMCP instead of brittle scraping or custom API integrations.
Secure, sandboxed V8 Isolate environments where agents can write and execute custom code with zero network access. Paired with @cloudflare/worker-bundler for runtime npm package installation. The agent can dynamically install libraries, compile them with esbuild on-the-fly, and run them in isolation.
Agent power tool
An agent that can write code, install packages, and execute them safely... all on Cloudflare's edge. No local Mac Mini needed. This is how you'd build an agent that can do data analysis, generate reports, or process files autonomously.
The playbook mentioned acquisitions through 2024 but did not cover these two 2026 acquisitions.
The playbook mentioned Workers run on Cloudflare's network. Gemini explains how at the architecture level:
The playbook mentioned Workers pricing. Gemini explains why it's so cheap with a concrete example:
Durable Objects automatically hibernate when WebSocket connections go quiet, saving state and freeing memory. The network connection stays alive at the routing layer. When a message arrives, the DO wakes up, restores state, and processes it. This enables millions of concurrent real-time connections at minimal cost.
Summary: What Gemini added
5 entirely new capabilities not in the playbook (Project Think, Browser Run Human-in-the-Loop, WebMCP, x402 payments, Code Sandboxes/codemode). 2 new acquisitions (Human Native, Astro). 3 areas with significantly deeper technical detail (V8 Isolates, CPU billing model, WebSocket hibernation). Plus architecture blueprints for 4 specific OA apps (see Blueprints tab).
Key details from Gemini's corporate research, including data points not in the playbook.
Founded July 26, 2009 by Matthew Prince, Lee Holloway, and Michelle Zatlyn. Origin: Project Honey Pot (spam/malware tracking). Grew from 5 engineers in Palo Alto to 5,156 employees by 2025.
NYSE as NET on September 13, 2019. $15/share, codenamed "Project Holloway." Dual-class voting: Class B shares carry 10 votes each, keeping Prince (~10% equity) and Zatlyn (~4% equity) in control. Institutional ownership at ~82% by Q1 2026 (led by Vanguard ~10.5%, BlackRock ~8.2%, Fidelity ~7.4%).
| Metric | Value |
|---|---|
| 2025 Revenue | $2.168B |
| 2025 Operating Income | -$207M |
| Q1 2026 Operating Cash Flow | $158.3M (up from $145.8M) |
| Q1 2026 Free Cash Flow | $84.1M (13% of revenue) |
| FY2026 Revenue Guidance | $2.805B - $2.813B |
| FY2026 Non-GAAP Operating Income | $418M - $421M |
Traditional clouds isolate apps using containers or VMs, each running a full OS. Cold starts: 100ms to 2 seconds. Cloudflare bypasses this entirely with V8 Isolates (Google's Chromium engine), running thousands of separate apps inside a single shared OS process. No OS boot. Sub-millisecond startup.
Extends Cloudflare's edge to the IP layer of enterprise networks. Ingests traffic at nearest data center, applies packet-by-packet DDoS mitigation, firewalling, and routing, delivers clean traffic back via GRE tunnels or private interconnects. Eliminates physical scrubbing appliances.
Mobile-first VPN using WireGuard protocol. WARP+ integrates Argo smart-routing backbone for dynamic congestion-aware routing through fastest physical paths.
A binding is a secure, pre-configured link that lets app code interact directly with platform resources (D1, R2, KV, Durable Objects) without external API keys, passwords, or connection management. For vibe coding: tell the AI to "use Wrangler bindings" instead of fetch() for clean data access.
Gemini's plain-English breakdown of every storage option, with pricing and limits.
| Product | Model | Consistency | Free Tier | Paid | Best For |
|---|---|---|---|---|---|
| Workers KV | Global key-value store | Eventual (60s propagation) | 100K requests/day | Incl. in $5/mo | Settings, configs, feature flags |
| R2 Storage | S3-compatible blob/object | Strong on writes/deletes | Generous limit | $0.015/GB/mo, $0 egress | Images, videos, PDFs, backups |
| D1 Database | Serverless SQLite | Strong on primary, eventual on replicas | Limited read/write rows | Incl. in $5/mo | Structured tables (users, orders) |
| Durable Objects | Stateful compute + SQLite | Linearizable strong (single-threaded) | Incl. in $5/mo | Usage-based | Real-time collab, chat, multiplayer |
| Hyperdrive | Connection pooler + edge cache | Inherited from target DB | Limited | Incl. in $5/mo | Speed up existing Postgres/MySQL |
| Vectorize | Semantic vector database | Index-specific | 30M queried dimensions/mo | $0.01/M queried | AI search, RAG, agent memory |
Gemini's recommendation for vibe coders
Use Drizzle ORM with D1 for type-safe SQL queries. Tell your AI assistant to use Wrangler bindings instead of standard fetch() for clean, robust data access code.
Gemini's deep dive into Cloudflare's AI agent architecture. This is the most strategically important section.
Workers AI runs 50+ open-source ML models on serverless GPUs at the edge. Pay-per-use. Models include LLMs (Llama, Mistral, Qwen, DeepSeek), image generation (Flux), embeddings (BGE), and translation.
AI Gateway is an observational proxy providing: query caching, rate limiting, automatic model fallbacks, and usage analytics. Works with Workers AI, OpenAI, Anthropic, Google, and Groq.
Manual protocol adapter. Developer handles: incoming chat messages, model stream mapping, text token blocks, tool executions. Best for simple, stateless chatbots without conversation memory.
Opinionated, stateful framework. Auto-manages: message history persistence (DO SQLite), tool-calling loops, client-side web tools, real-time stream resumption. Capability-based authorization (agent starts with zero access).
Headless Chrome on Cloudflare's network. The agent writes JavaScript CDP commands against the browser window. Key features:
Secure V8 Isolate environments where agents write and execute code with zero network access. @cloudflare/worker-bundler enables runtime npm package installation, compiled on-the-fly with esbuild. The Tier 4 Sandbox provides a full secure OS container for cloning repos, compiling code, and running tests.
WebMCP: Open standard for websites to declare their APIs and actions for AI agents. Agents discover and interact with sites through structured endpoints instead of scraping.
x402: Machine-to-machine payment protocol (with Coinbase). Agents use micropayments to buy API access, datasets, browser rendering hours, or search queries on the fly. Processed via Stripe or stablecoin networks.
The full agent architecture (from Gemini's diagram)
Inbound Event (Chat, Voice, Webhooks, Email, Slack) → Think Agent Harness (extends Agent Class, manages AI Chat Protocol, runs inside Durable Object) → DO SQLite Store (conversation history, state) → Workers AI (model execution on edge GPUs, tracked via AI Gateway) → Agent decides to call external tools → Browser Run (headless Chrome, CDP commands, human-in-the-loop) or Code Sandbox (secure container, dynamic code execution, isolated workspace).
Gemini designed specific Cloudflare architectures for four of Brad's apps. These are reference patterns we can build from.
UX: Single search bar + 3 dynamic columns (Immediate Focus, Awaiting Input, Backlog). Auto-updates without page refreshes.
Architecture:
vs. what we built
We already have Command Center on Pages + D1 + Workers Functions. What Gemini adds: Vectorize for semantic search across emails/contacts (we don't have this), Email Worker for automatic email ingestion (we poll via Gmail API), and Durable Objects for real-time push updates (we use manual refresh/polling). All three are legitimate upgrades.
UX: Clients log in to view project tasks, download reports, interact with team workflows. Managers assign tasks, track hours, upload deliverables.
Architecture:
vs. what we have
server.outsourceaccess.com is the current OA portal. Client dashboards are separate Pages projects with custom auth tokens. Gemini's blueprint uses Access (proper Zero Trust) instead of our custom token approach, and Workflows for reliable multi-step processes. Both are improvements worth considering as the portal evolves.
UX: Coach on a tablet drags/drops player icons. Other coaches see updates in real time. Parent portals update instantly.
Architecture:
vs. what we built
DugoutReady exists as a WordPress site (dugoutready.com on Cloudways). The Cloudflare-native version with Durable Objects for real-time sync would be a significant upgrade. This is the architecture for the full product build.
UX: Shared tablet dashboard in the kitchen. Family calendar, dynamic chore lists, document upload portal, AI chat assistant.
Architecture:
vs. what we built
Stevens Family Planner exists on Pages + D1 + Workers. The Gemini blueprint adds: R2 for document storage (we don't have this yet), Workers AI + Vectorize for "chat with your family documents" (we don't have this). Both are strong feature additions.
This aligns with the playbook's findings but adds the specific data flow guardrails. Both sources agree: get written confirmation from Cloudflare sales before building any PHI-handling app on the developer platform.
Gemini's analysis of WHY Cloudflare is so cheap, not just what the prices are.
Traditional clouds allocate dedicated CPU + RAM per container. Resources sit idle between requests. Idle cost is baked into pricing. V8 Isolates share physical resources across thousands of apps. No idle cost to pass on.
Traditional: billed for total elapsed duration (including idle waiting for API responses). Cloudflare: billed only for active CPU milliseconds. A Worker waiting 10s for an API but using 10ms CPU = billed for 10ms. Massive savings for AI agents.
Traditional: keeping WebSocket alive = keeping container running in memory. Cloudflare: Durable Object hibernates when connection goes quiet, frees memory. Network connection stays alive at routing layer. DO wakes on new message. Millions of concurrent connections at minimal cost.
R2's zero egress fees. Cloudflare's existing global network for security means serving storage/compute costs almost nothing extra. 10 TB of files monthly: AWS S3 egress ~$900. R2 egress: $0.
| Dimension | Cloudflare | Vercel | Netlify | Supabase | Railway |
|---|---|---|---|---|---|
| Compute Model | V8 Isolates | AWS Lambda | AWS Lambda | Dedicated Postgres | Docker Containers |
| Cold Start | <1ms | 100ms-2s | 100ms-2s | Connection pool dependent | Container dependent |
| Egress Fees | Zero | High markups | High markups | Standard cloud | Standard cloud |
| Stateful Mechanics | Durable Objects + D1 | Third-party | Third-party | Real-time Postgres | Persistent volumes |
| AI Native | Workers AI, Vectorize, Think | Third-party wrappers | Third-party wrappers | pgvector | General hosting |