Cloudflare Project Think is an agents software development kit (SDK) designed to enable the creation of next-generation AI agents with advanced capabilities for durable execution and complex task orchestration. The platform represents Cloudflare's strategic initiative to provide infrastructure and tooling that positions intelligent agents as compositions of multiple complementary systems including runtime environments, user interface grounding, browser automation, voice interfaces, and sandboxed code execution capabilities.

Project Think centers on a durable execution model that ensures agent operations persist and recover reliably even in the face of system failures or interruptions. This architectural approach addresses a critical limitation in many AI agent implementations: the inability to maintain consistent state and execution continuity across extended operations. The SDK provides native support for sub-agents, enabling hierarchical agent structures where parent agents can delegate tasks to specialized child agents, facilitating more modular and scalable agent composition patterns ¹⁾.

The platform includes persistent sessions functionality that allows agents to maintain context and state information across multiple interaction cycles. This capability is essential for agents that require memory of previous interactions, user preferences, and evolving task contexts over time. By providing persistent session management at the infrastructure level, Project Think reduces the complexity developers face when implementing stateful agent behaviors.

Sandboxed Code Execution represents a critical security and isolation mechanism within Project Think. The SDK provides isolated execution environments where agent-generated or agent-executed code runs within controlled boundaries, preventing unauthorized access to host systems or unintended side effects on the broader infrastructure. This approach aligns with industry best practices for safely deploying autonomous agents in production environments. Project Think specifically includes Agent Lee, an in-dashboard agent with sandboxed TypeScript execution capabilities designed for UI automation tasks ²⁾.

The built-in workspace filesystem provides agents with a dedicated file system abstraction for managing data, storing intermediate results, and organizing computational artifacts. This filesystem abstraction enables agents to perform file-based operations—reading, writing, and manipulating data structures—without requiring direct access to underlying infrastructure systems.

Runtime Tool Creation allows agents to dynamically generate and instantiate new tools during execution rather than being limited to a pre-defined set of capabilities at deployment time. This dynamic capability expansion enables agents to adapt their instrumental capabilities to specific task requirements and evolving problem contexts ³⁾

Project Think functions as a component within Cloudflare's comprehensive vision of agent architecture, which integrates multiple modalities and interaction paradigms. The platform conceptualizes agents as compositions of durable runtime infrastructure (the execution and state management layer), UI grounding (connecting agent outputs to user-facing interfaces), browser automation (enabling web-based task execution), voice interfaces (supporting conversational interaction modalities), and sandboxed execution environments (ensuring secure code execution). Cloudflare has complemented Project Think's core SDK with additional production-ready components, including Agent Lee, an in-dashboard agent designed for UI automation tasks, and real-time voice pipelines delivered over WebSockets to enable voice-based agent interactions ⁴⁾. The company has also rebranded its Browser Run stack with live view and session recording capabilities to enhance browser automation and observability for agent-based workflows ⁵⁾.

This compositional approach recognizes that sophisticated agent systems require integration across multiple technological domains. Rather than treating agents as monolithic entities, the architecture emphasizes modularity and integration of specialized subsystems, each optimized for specific functional requirements. Browser automation capabilities enable agents to interact with web applications and services. Voice interfaces extend agent accessibility to conversational modalities beyond text. UI grounding mechanisms ensure that agent-generated outputs map appropriately to user-facing displays and interaction paradigms.

As part of Cloudflare's broader Workers ecosystem—the company's serverless computing platform—Project Think leverages existing infrastructure for edge computing, distributed execution, and global deployment. This integration positions agent workloads to run proximate to end users and data sources, reducing latency and improving responsiveness. The durable execution capabilities align with Cloudflare's Durable Objects technology, which provides strongly consistent, globally coordinated execution primitives suitable for maintaining agent state across distributed environments ⁶⁾.

Project Think reflects broader industry trends toward making agent development more accessible through purpose-built infrastructure and development frameworks. The emphasis on durable execution, security through sandboxing, and modular composition addresses key technical challenges that organizations face when deploying AI agents in production environments. By providing these capabilities as part of a unified SDK rather than requiring developers to assemble them from disparate systems, Cloudflare aims to accelerate agent adoption and reduce time-to-market for agent-based solutions.

• Cloudflare Agent Lee
• Goose
• OpenAI Agents SDK
• How to Create an Agent
• Vercel AI SDK