Cloudflare absolutely crushed its five-day Agents Week sprint

Cloudflare capped its five-day Agents Week sprint on August 7, 2026, with more than two dozen announcements aimed at positioning itself as the operating layer for software agents, integrating its Workers, Zero Trust, AI Gateway, and network into a unified Agent Cloud pitch. The company reported $2.168 billion in 2025 revenue and approximately 332,000 paying customers as of December 31, 2025, but did not disclose agent-specific usage or revenue figures. Key releases included an early preview of the @cloudflare/computer runtime, cross-language Workers RPC, and a Billable Usage API.

Cloudflare https://cloudflare.com/ , the Internet infrastructure provider co-founded by Matthew Prince @eastdakota https://x.com/eastdakota and Michelle Zatlyn @zatlyn https://x.com/zatlyn , capped a five-day product push on Monday with a clear claim: Cloudflare wants to supply the operating layer for software agents, from the computer that executes a task to the identity, payment and access systems governing what happens next. The Agents Week recap https://blog.cloudflare.com/agents-week-review-august-2026/ collected more than two dozen announcements Cloudflare published from August 3rd through August 7th. The volume can obscure the strategy. Cloudflare is joining products it already sells across Workers, Zero Trust, AI Gateway and its global network into a single Agent Cloud pitch. That gives Cloudflare an answer to hyperscalers selling model access and compute, while directing agent workloads through services Cloudflare can observe, secure and bill. Prince, Zatlyn and Lee Holloway started Cloudflare after Project Honey Pot users asked for a way to stop abusive Internet traffic instead of merely tracking it. The founders have repeated the same expansion pattern since: turn network visibility into controls, then expose those controls as a platform. Agents Week applied that pattern to autonomous software, where the requests can trigger code deployments, purchases and changes inside corporate systems. Cloudflare entered the push with substantial distribution. It reported $2.168 billion in 2025 revenue and approximately 332,000 paying customers as of December 31st, 2025. Agents Week did not include customer counts, revenue or usage figures for the agent products, leaving their commercial contribution unclear. Monday: giving each agent a computer Cloudflare opened the week with an early preview of @cloudflare/computer , a runtime that presents an agent with a filesystem, shell and tools while deciding whether work should run inside a lightweight isolate, a container sandbox or a browser. RuntimeWire covered the architecture when it launched /article/cloudflare-computer-agent-runtime-isolates-containers : routine operations can stay in fast isolates, while heavier Linux workloads move into containers. That abstraction is central to Cloudflare's argument. Developers should specify what the agent needs to do rather than manually provision and coordinate every execution environment. The package remains an early preview, so the announcement establishes the programming model before proving how it performs under large production workloads. Cloudflare also made Workers RPC cross-language https://blog.cloudflare.com/python-workers-rpc/ , allowing Python and JavaScript Workers to call each other's methods and share objects without developers defining a separate schema or serialization layer. The feature extends Cloudflare's Cap'n Proto-based RPC system and reduces the plumbing required when an agent application combines Python model code with TypeScript services. A separate engineering post described how Cloudflare is serving Kimi and GLM models https://blog.cloudflare.com/smaller-faster-safer-models/ using KV-cache quantization, compressed weights and cache protections. Cloudflare says those changes let it place more requests on shared GPUs without changing model accuracy. The work runs on the open-source SGLang inference framework. This was an infrastructure explanation rather than a new model release, aimed at showing Cloudflare can operate demanding open models on its own GPU fleet. The Billable Usage API https://blog.cloudflare.com/billable-usage-api/ gives self-serve accounts one endpoint for usage and cost data across Workers, R2, D1, Workers AI, Vectorize, Images and Stream. Cost reporting is a practical prerequisite for autonomous provisioning: a human-readable dashboard becomes inadequate when software is creating resources and spending continuously. Cloudflare rounded out Monday with inbound TCP and gRPC support https://blog.cloudflare.com/grpc-workers/ for Workers and Containers. A new connect handler accepts sockets from Cloudflare Spectrum, while Containers can run full-duplex gRPC servers. Workers can also serve unary and server-streaming gRPC APIs through gRPC-web translation. The features are in private beta, with voice agents and other low-latency applications presented as the initial use cases. Tuesday: Cloudflare tries to own the development loop Tuesday introduced Cloudflare's Agent Development Lifecycle https://blog.cloudflare.com/agent-development-lifecycle/ , or ADLC, a framework for managing software production when code generation becomes cheap and review, testing, deployment and maintenance become the bottlenecks. The term is marketing, but the operational problem is real: faster code output increases pressure on every control surrounding implementation. The first product under that framing, Cloudflare Agents https://blog.cloudflare.com/agents-on-cloudflare/ , starts with agent-aware tracing. It records model calls, tool executions, tokens and cost, with initial support for OpenTelemetry-compatible harnesses including Think, Flue and AI SDK. Cloudflare's broader plan includes replay and approval controls, though the first release concentrates on observability. Cloudflare moved some of that feedback into local development through automatic tracing in https://blog.cloudflare.com/local-tracing/ wrangler dev and vite dev . Supported coding agents can discover a local API, inspect OpenTelemetry traces, diagnose a failing Worker and verify a fix before deployment. The API exposes an OpenAPI schema at runtime, avoiding agent-specific setup instructions. Cloudflare Wallets https://blog.cloudflare.com/wallets/ was Tuesday's widest commercial bet and its least complete product. Users can reserve a wallet handle that Cloudflare intends to use as an agent identity for buying API access and content. Actual payments, Virtual Wallets and funding tools have no established release date. RuntimeWire reported on the gap between the available handle reservation and the planned transaction system /article/cloudflare-wallet-handles-ai-agent-spending . Cloudflare's CI SDK https://blog.cloudflare.com/ci-workflows/ connects its Artifacts code storage with Workflows, isolated builds, dependency caching, tests and conditional deployments. An optional review agent can inspect a broken pipeline and prepare a corrective commit. Cloudflare is pitching the system to platforms that manage large numbers of customer repositories and want to define the build process once instead of forcing each customer to operate separate CI infrastructure. Two internal case studies supplied evidence for the lifecycle pitch. Cloudflare said its AI code reviewer https://blog.cloudflare.com/engineering-standards-enforcement/ flagged nearly 250,000 deviations from engineering standards and blocked 16,000 merges during a four-month period. A specification reviewer evaluated close to 600 technical designs. Both use the Cloudflare Codex, a governed collection of internal standards organized by engineering domain. The second case study described an automated triage pipeline for Astro https://blog.cloudflare.com/astro-issue-triage/ that reads reports, reproduces bugs in sandboxes and produces preview releases. Cloudflare said the pipeline reduced open issues from more than 200 to about 30. Despite the post's headline, the count had not reached zero when Cloudflare published it; the team said it expected to reach that point within a month. The underlying framework, Flue, is open source. Wednesday: identity and access move from people to tasks Cloudflare's Agent Access Model https://blog.cloudflare.com/the-agent-access-model/ treats each agent run as a short-lived principal with its own capability ceiling and trust level. Traditional Zero Trust systems authenticate a human and inspect a device. Agents can spawn multiple tasks, operate much faster and continue acting under credentials that reveal little about the specific job being performed. The proposed model narrows each task's permissions instead of relying on a single risk decision at the start of a session. That approach fits Cloudflare's security business: if agents become a common way to reach internal applications, Cloudflare can extend Access, Gateway and its policy engine to a new class of requester. Cloudflare also open-sourced Cloudflare OS https://blog.cloudflare.com/cloudflare-os/ , an internal workspace that gives employees an agent, company context and access to approved systems. Cloudflare says it distributed the first version to every employee in May and that thousands of employees use it daily. The open-source version can be deployed by other organizations and connected to their internal services. Cloudflare CIO Sam Rhea said the project began after employees started requesting broad API access for agent-built internal applications. Cloudflare assembled the initial system from Workers and Access, then added services for shared context, permissions and collaboration. The origin matters because Cloudflare is using its own adoption problem to shape a product it can sell alongside Zero Trust. Identity-aware AI Gateway https://blog.cloudflare.com/identity-aware-ai-gateway/ entered open beta, while User Insights became generally available at no additional charge to AI Gateway customers. Requests authenticated through Cloudflare Access receive a verified user identifier, allowing administrators to analyze model usage and spending by person or agent. Spending rules can block requests or move a user to a cheaper model after a budget threshold. RuntimeWire examined the 30-day behavioral baseline and anomaly controls /article/cloudflare-ai-gateway-identity-rogue-agent-alerts . WriteGuard https://blog.cloudflare.com/mcp-portal-writeguard-private-beta/ adds centralized controls for write operations made through Model Context Protocol servers. Cloudflare built it after expanding its internal MCP portal from 13 servers in April to 27 by August. The controls can distinguish agent activity from a user's manual actions, restrict dangerous tools and preserve an audit trail. WriteGuard is in private beta, limiting the immediate availability of one of the week's most important security products. Thursday: rebuilding websites for machine visitors Cloudflare's Agentic Internet proposal https://blog.cloudflare.com/the-agentic-internet/ divides the machine-readable web into four requirements: content must be readable, discoverable, callable and payable. Cloudflare argues that agents repeatedly downloading unchanged pages wastes compute and forces websites built for visual navigation into an inefficient interface for software. That framing puts Cloudflare between agents and website owners. Cloudflare can identify the agent, render or restructure the site, expose callable tools, enforce access rules and eventually process payment. Each added function increases the value of routing both sides through Cloudflare's network. WebMCP on Cloudflare https://blog.cloudflare.com/webmcp/ is a developer preview that can add machine-callable tools to a website without changing its origin code. Cloudflare injects a bridge that registers tools through the experimental document.modelContext interface in Chrome 146. Browser agents can call those tools directly instead of interpreting buttons and forms through screenshots or page markup. Cloudflare paired WebMCP with Agent Readiness and Answer Engine Optimization tools https://blog.cloudflare.com/aeo/ in its dashboard. Agent Readiness shows how software sees and uses a site, while the AEO product tracks whether answer engines recommend it. Cloudflare says fewer than half of HTML page requests come from humans, though that measurement includes many automated requests that are not personal agents. Kitesurf https://blog.cloudflare.com/kitesurf/ is Cloudflare's Rust-based browser for agents, compiled to WebAssembly and designed to run inside V8 isolates on Workers. It avoids some of the cost of full Chromium by sacrificing pixel-perfect rendering and portions of web compatibility. RuntimeWire reported that Cloudflare's tests showed lower CPU and memory use alongside slower page loads and narrower compatibility /article/cloudflare-kitesurf-browser-ai-agents-workers-2 . It remains a beta and is best suited to tasks where extracting and acting on structured page content matters more than visual fidelity. Cloudflare also adapted to the July 28th MCP specification https://blog.cloudflare.com/mcp-v2/ , which rewrote the protocol around stateless requests and updated its TypeScript, Python, Go and C SDKs. Stateless MCP servers can run directly in Workers without sticky sessions, open streams or separate stateful coordination. Durable Objects remain available where an application needs persistent state or human interaction. AI Search https://blog.cloudflare.com/ai-search-easier/ now packages crawling, storage, embeddings, vector search and reranking that previously required developers to assemble Workers AI, AI Gateway, Vectorize, R2 and Browser Run. It can index files or verified websites, expose /search and /mcp endpoints, operate behind a custom domain and add semantic search to sites built with Cloudflare's EmDash CMS. Cloudflare also published an early pricing preview in which embeddings and reranking are included when customers use selected default models. Friday: Cloudflare consolidates the control plane Friday's bot behavior paper https://blog.cloudflare.com/good-and-bad-agentic-behaviors/ argued for evaluating risk and trust continuously as sessions shift between human and automated activity. A shopper might browse manually and then hand checkout to an agent, making a static human-or-bot label less useful. Much of the post described Cloudflare's product direction rather than shipping a discrete new control. Cloudflare made a more concrete consolidation move by joining Workers AI and AI Gateway https://blog.cloudflare.com/workers-ai-gateway-unification/ behind one binding, REST endpoint, wallet and dashboard. Workers AI hosts models on Cloudflare's GPUs; AI Gateway proxies requests to Cloudflare and third-party providers while applying logging, security and spending controls. A shared interface removes an early architectural choice for developers and prepares Cloudflare for model-first routing across providers. Some of that unification is available, while broader routing plans remain future work. The community announcement created Cloudflare Ambassadors and Community Engineers https://blog.cloudflare.com/community-program-refresh/ and committed another $1 million to open-source funding over two years. Ambassadors will organize and teach in local communities. Community Engineers will work on open-source infrastructure. The program also serves Cloudflare's distribution strategy by putting trained advocates and maintainers around a developer platform that is broadening quickly. Finally, Radar Researcher https://blog.cloudflare.com/introducing-radar-researcher/ entered beta. The assistant converts plain-English questions into queries over Cloudflare Radar data and returns interactive charts. It uses Radar's API rather than asking a model to generate unsupported measurements. RuntimeWire covered how product lead Lai Yi Ohlsen designed the feature around an auditable data trace /article/cloudflare-radar-researcher-internet-data-ai-charts . One stack, several maturity levels Agents Week gave Cloudflare a coherent stack on paper: @cloudflare/computer supplies execution, Cloudflare Agents and local tracing observe it, the ADLC manages changes, Access and WriteGuard constrain actions, Wallets handles identity and planned payments, and WebMCP, Kitesurf and AI Search connect agents to the web and private data. The releases are unevenly mature. Cross-language RPC, the Billable Usage API and local tracing are usable additions to existing products. Identity-aware AI Gateway is in open beta. Cloudflare Computer is an early preview. Inbound TCP, gRPC and WriteGuard require private-beta access. WebMCP is a developer preview built around an experimental browser standard. Kitesurf and Radar Researcher are betas. Wallets currently offers handles while the payment system remains planned. Cloudflare has still assembled a broader agent infrastructure position than a model-hosting announcement alone would provide. Prince and Zatlyn built Cloudflare by placing security, performance and developer services on one network. Their next platform bet assumes autonomous software will need the same consolidation, with tighter controls because agents can spend money, deploy code and alter systems at machine speed.