Cloudflare has introduced the Cloudflare Agents SDK to provide a foundational platform for deploying production-grade AI agents. This platform addresses common distributed systems challenges, such as automatic state resumption, secure untrusted code execution, and persistent filesystem access. The SDK is designed to support various agent harnesses—the software layers managing an agent's loop and tool access—by offering primitives like durable execution and dynamic workflows.

Flue, a new open-source framework, is the first to build upon this SDK. Shipped in 1.0 Beta on June 17, 2026, Flue allows developers to build agents using a declarative model where they describe the agent's context and instructions rather than scripting every step. It includes native frontend hooks via @flue/react for streaming agent state and tool execution directly into user interfaces, and provides pre-configured integrations for platforms like Slack, GitHub, and Discord.

To ensure reliability, the platform utilizes Durable Objects and a checkpointing mechanism called runFiber(). This allows agents to recover from interruptions by processing execution history through an append-only log, ensuring state is never lost. Additionally, the SDK includes @cloudflare/codemode for executing LLM-generated code in isolated workers, which start in under 10ms at a cost of $0.002 per load. For filesystem tasks, @cloudflare/shell provides a virtual, SQLite-backed filesystem that avoids the overhead of full containers.

The SDK also integrates with the broader Cloudflare ecosystem, granting agents secure access to capabilities like web browsing, email services, and persistent memory without exposing credentials in agent-generated code. Agents running on this platform can also trigger multi-step, durable workflows using the Workflows engine, allowing for consistent task execution even across long durations or external dependencies. This architecture mirrors the foundations used in Project Think, Cloudflare's first-party agent harness, now made available for the open-source community.

Cloudflare has introduced the Cloudflare Agents SDK to provide a foundational platform for deploying production-grade AI agents. This platform addresses common distributed systems challenges, such as automatic state resumption, secure untrusted code execution, and persistent filesystem access. The SDK is designed to support various agent harnesses—the software layers managing an agent's loop and tool access—by offering primitives like durable execution and dynamic workflows.

Flue, a new open-source framework, is the first to build upon this SDK. Shipped in 1.0 Beta on June 17, 2026, Flue allows developers to build agents using a declarative model where they describe the agent's context and instructions rather than scripting every step. It includes native frontend hooks via @flue/react for streaming agent state and tool execution directly into user interfaces, and provides pre-configured integrations for platforms like Slack, GitHub, and Discord.

To ensure reliability, the platform utilizes Durable Objects and a checkpointing mechanism called runFiber(). This allows agents to recover from interruptions by processing execution history through an append-only log, ensuring state is never lost. Additionally, the SDK includes @cloudflare/codemode for executing LLM-generated code in isolated workers, which start in under 10ms at a cost of $0.002 per load. For filesystem tasks, @cloudflare/shell provides a virtual, SQLite-backed filesystem that avoids the overhead of full containers.

The SDK also integrates with the broader Cloudflare ecosystem, granting agents secure access to capabilities like web browsing, email services, and persistent memory without exposing credentials in agent-generated code. Agents running on this platform can also trigger multi-step, durable workflows using the Workflows engine, allowing for consistent task execution even across long durations or external dependencies. This architecture mirrors the foundations used in Project Think, Cloudflare's first-party agent harness, now made available for the open-source community.