Cloudflare put Workers at the center of its Agent Cloud push

Why it matters

Cloudflare is positioning its network as the control layer between agents and the web, extending its security and developer businesses into machine-run workloads.

In April 2026, Cloudflare Vice President of Product Rita Kozlov used the welcome post for Cloudflare's first Agents Week to recast Workers as infrastructure for autonomous software. The pitch covered the full operating environment an agent needs: compute, persistent state, identity, private access, search, memory, browser control and a path onto the open web.

Cloudflare credited the welcome post to Kozlov. Her Cloudflare TV profile says she joined Cloudflare in 2016 and spent much of the next decade helping expand Workers from an edge-compute service into a broader developer platform, putting her close to the platform history behind the Agent Cloud push.

Cloudflare's founders had followed a similar path before. Matthew Prince, Michelle Zatlyn and Lee Holloway built Cloudflare after Prince and Holloway's Project Honey Pot attracted users who wanted help stopping abusive traffic, rather than merely identifying its source. Cloudflare subsequently combined security and performance on the same network. The Agent Cloud strategy applies that playbook to a new workload, bundling execution with the controls needed to let software act on private data and external services.

Prince made the ambition explicit in Cloudflare's April 13 press release, saying agents needed a secure home that could persist across long-running tasks. He described Workers, which Cloudflare had spent nine years building at that point, as the foundation for making Cloudflare the platform for an agent-driven web.

One stack assembled from many products

Kozlov's product move assembled existing Cloudflare layers and new services into one stack, rather than asking developers to adopt a separate agent platform. Workers supplied lightweight execution. Durable Objects and Workflows supplied state and durable processes. Cloudflare Access and Gateway supplied policy enforcement. The global network supplied distribution.

During the week, Cloudflare's April 20 review listed Git-compatible Artifacts storage, generally available Sandboxes, programmable egress controls and isolated SQLite databases for applications created by Dynamic Workers. Cloudflare said a redesigned Workflows control plane could support 50,000 concurrent processes, a Cloudflare-reported capacity figure rather than a measure of customer demand.

The security releases addressed the harder operational problem: agents need credentials and access to internal systems if they are expected to complete useful work. Cloudflare Mesh connected agents to private networks, while Managed OAuth for Access was designed to let an agent act with a user's authorization. Cloudflare also added controls for non-human identities and published an architecture for governing Model Context Protocol servers.

The toolbox layer included Agent Memory, AI Search and Browser Run. Cloudflare also said its inference layer could call models from more than 14 providers. That model-neutral position is central to the strategy. Cloudflare does not need to win the contest to train the most capable model if it can become the network, runtime and security layer through which different models take action.

The hyperscalers are building the same control points

Cloudflare is competing for a workload that Amazon, Microsoft and Google already treat as an extension of their cloud platforms. Amazon Bedrock AgentCore grouped runtime, memory, identity, browser access, code execution and observability into modular services. Microsoft Foundry Agent Service paired an agent runtime with private networking, OAuth, evaluations and Azure integrations. Google has been adding managed agent deployment and related services through Vertex AI.

That overlap validates Kozlov's premise while raising the bar for differentiation. Memory, sandboxes and tool gateways are becoming standard cloud features. Cloudflare's advantage rests on the network it already operates and the ability to enforce policies close to users, applications and data across different cloud providers.

Cloudflare says its network reaches 335+ cities across 125+ countries and handles roughly 20% of web traffic. Its 2025 annual filing reported approximately 332,000 paying customers and 4,298 large customers as of December 31. That distribution gives Cloudflare a large base into which it can sell agent infrastructure without first convincing customers to move their primary databases or model training workloads.

The commercial test is still ahead

Cloudflare reported $639.8 million in first-quarter 2026 revenue, up 34% from the prior year. Agent workloads can increase demand for compute, storage and network services, though they also introduce infrastructure costs that Cloudflare will have to manage carefully.

Cloudflare's disclosures do not separate Agent Cloud revenue, paid adoption or customer counts from its broader developer and security businesses. The commercial test is whether agents create additional usage across Workers, storage and Zero Trust products, rather than shifting workloads among services customers already buy.

The Agents Week framing also pushed beyond hosting agent code. Cloudflare's materials put discovery, access and payments on the agentic web agenda, alongside controls for agents reaching systems of record. That extended the strategy from agent execution into the economics of how autonomous software discovers and consumes information.

Kozlov's Agent Cloud thesis gives Cloudflare a coherent route into AI infrastructure without requiring Cloudflare to become a frontier model laboratory. Cloudflare can instead own the boundaries agents repeatedly cross: between code and execution, identity and permission, private systems and the public web, and content access and payment. The breadth of those control points is the bet. Adoption and revenue will determine whether they become a durable business.