Cloudflare Agents SDK brings durable execution and state management to production agents; Flue framework 1.0 ships

Cloudflare released the Agents SDK, a base-layer runtime and platform for building production-grade agentic AI systems. The SDK provides three core primitives: durable execution (appending every prompt, tool call, and model response to an immutable append-only log), a durable filesystem (agents can read/write artifacts and code without losing state on crashes), and dynamic workflows (sandboxed code execution via @cloudflare/codemode). These primitives let any agentic harness (Claude Code, Pi, Project Think, OpenCode) run reliably in the cloud without rewriting orchestration logic.

Flue, the first framework built on the Agents SDK, shipped 1.0 Beta this week. Built on the Pi harness (the same foundation as OpenClaw), Flue uses a declarative model: you define the context an agent needs (model, skills, instructions, sandbox) and it solves the task autonomously—no orchestration loop required. Flue integrates directly with Slack, GitHub, Linear, Discord, and custom frontends via native React hooks. The framework also supports multi-cloud deployment: agents run as long-lived Node.js processes on any VM, GitHub Actions, or Cloudflare Durable Objects (for automatic scaling and isolated storage per agent).

Cloudflare's design philosophy separates three layers: (1) the framework (Flue) for developer experience and integrations; (2) the harness (Pi/Project Think) for agentic loops and tool use; (3) the platform (Agents SDK) for compute, state, and storage. When agents deploy to Cloudflare, each runs inside its own Durable Object, avoiding manual server provisioning, sticky sessions, and noisy-neighbor problems. Durable Streams ensure agents survive crashes: if a process dies, another picks up the append-only log and resumes from the exact step it left off.

For architects: this is the first major platform offering that treats distributed systems failures (crashes, timeouts, lost context) as a solved problem for agents in production. The Agents SDK is framework-agnostic—any harness can adopt it—which lowers the barrier for porting agents from local development to cloud-native workloads. Practitioners should evaluate: (1) whether durable execution model maps to your agent use case (stateful multi-step tasks benefit most); (2) integration density with your existing tools (Slack/GitHub depth is strong); (3) pricing model for Durable Objects at scale. Early adopters in customer triage, code review, and incident response workflows are already live.