Rita Kozlov

Rita Kozlov is Cloudflare’s vice president of product for its developer platform and AI, shaping infrastructure that helps developers build globally distributed applications and autonomous agents. Her work spans Cloudflare Workers, Durable Objects, and the execution, security, and coordination systems that make agentic software dependable.

Kozlov studied computer science at Georgia Tech and interned at Microsoft, where she worked on the Windows Phone software development kit. She became a software engineer at 6sense before joining Cloudflare in 2016 as a solutions engineer serving enterprise customers. She subsequently moved into product management for Workers and advanced into broader leadership of Cloudflare’s developer platform and AI. Her account of that transition traces a career grounded in software development, customer needs, and product design.

Using Workers internally exposed practical obstacles in local development and debugging, informing improvements to developer tooling. Kozlov treats building with the company’s own platform as a direct way to understand which infrastructure decisions genuinely simplify developers’ work.

How she thinks about dependable agents

• Agents require complete application architecture. Kozlov divides agentic systems into four cooperating components: clients, model reasoning, workflows, and tools. A CRM assistant might combine voice or chat, customer records, model inference, outbound communications, and human review; the model supplies reasoning, while surrounding infrastructure handles execution.

• Persistent state prevents operational failures. Durable Objects keep application logic and state together, allowing agents to remember preferences across sessions, recover after interruptions, and resume delayed tasks. In a card-approval workflow developed with notification provider Knock, persistent state routes approvals to the correct agent and prevents duplicate card issuance.

• Tool access needs security and coordination. The Model Context Protocol lets multiple clients connect to remote tools, but production services still need authentication, streaming connections, and durable memory. Kozlov positions the Cloudflare Agents SDK and its McpAgent class as reusable infrastructure for those requirements, illustrated in her AI Engineer World’s Fair talk.

• Human approval belongs inside execution. Her human-in-the-loop workflows pause consequential actions until an authorized person responds, then resume from recorded state. Approval is part of the application’s control flow, not an informal safeguard layered over autonomous behavior.

In November 2025, Kozlov and Replicate co-founder Ben Firshman announced Replicate’s integration into Cloudflare, bringing AI-model deployment closer to Workers and its surrounding developer platform. Her subsequent writing outlines an agent-native cloud built around durable execution, storage, identity, security, and access to existing online services.