AIAny
AI Agent2025
Icon for item

Cloudflare Agents

Runs stateful AI agents as Cloudflare Durable Objects — each keeps its own storage and lifecycle, hibernating when idle and waking on demand. Adds WebSocket state sync, type-safe RPC, resumable LLM streaming, MCP roles, and durable workflows.

Introduction

Most agent frameworks treat persistence as something you bolt on afterward — a Redis instance for memory, a separate session store, a queue for long jobs. Cloudflare Agents collapses all of that into one primitive: the agent is a Durable Object. Compute, storage, and identity live in the same place, so an agent can sleep for days, wake on a single request, and resume exactly where it left off without rehydrating state from anywhere else.

What Sets It Apart
  • State and compute are co-located. Each agent owns embedded SQLite storage that survives restarts and syncs to connected clients automatically — no external database round-trips on the hot path.
  • Hibernation is the default, not an optimization. Idle agents cost nothing and evict from memory; an incoming WebSocket message or RPC call revives them, making "one long-lived agent per user, session, or document" economically viable at scale.
  • Resumable LLM streaming survives disconnects. Because the agent holds the stream's state, a client that drops mid-generation can reconnect and keep reading — a hard problem in stateless serverless setups.
  • MCP works both ways. An agent can act as an MCP server exposing its tools or as a client consuming external ones, and it can speak WebSocket, HTTP, email, and voice (STT/TTS) from the same object.
Who It's For

Great fit if you already build on Cloudflare Workers and want stateful, per-entity agents — chat assistants, document collaborators, long-running workflow runners — without standing up your own state layer. The Durable Object model hands you strong single-threaded consistency per agent for free.

Look elsewhere if you need to run on your own hardware or another cloud: the framework is tightly coupled to Durable Objects and Workers, so it isn't portable. Likewise, if your agent is a one-shot stateless call, the persistence machinery is overhead you won't use.

More Items

Enables RL post-training with million-token prompts under a fixed GPU budget by evaluating shared prompt state without autograd, retaining only minimal model state, and replaying short response branches; instantiated as GRPO and demonstrated on Qwen3.6-27B and GLM-5.2 up to multi-million token execution.

GitHub
AI Infra2026

Defines OpenTelemetry semantic conventions for generative AI telemetry — spans, metrics, and events for GenAI clients, the Model Context Protocol (MCP), and provider-specific integrations. Includes YAML models, human-readable docs, and reference implementations to standardize observability across GenAI deployments.

GitHub
AI Agent2025

Autonomous Red Team agent that plans and executes realistic attack chains (reconnaissance, exploitation, pivoting, C2) while producing Rules of Engagement and an OPPLAN and running actions inside a hardened sandbox; intended for authorized red-team engagements and defensive verification.