Project Think: Building Next-Generation AI Agents on Cloudflare

Project Think extends Cloudflare's Agents SDK a step further, bundling core primitives for long-running AI agents. The goal is not a simple chatbot, but agent infrastructure that continues without interruption, collaborates, and uses tools on its own.

Existing agents are tied to laptops or expensive VPS, cost money even when idle, and are cumbersome to install, update, and manage secrets for. Cloudflare sees AI agents as fundamentally 1:1 in structure, requiring a different scaling model from traditional 1:N applications.

The core idea is to run agents on top of Durable Objects, giving each agent a unique identity, state, and SQLite. When hibernated, an agent sleeps at 0 compute cost, and wakes up to restore state and continue work when a request, WebSocket, schedule, or email event arrives.

For long-running tasks, it provides durable execution with fibers. You register execution with runFiber(), leave intermediate checkpoints with stash(), and recover on restart with onFiberRecovered. When needed, keepAlive() and keepAliveWhile() prevent eviction during short tasks.

Work is also separated via sub-agents. A parent agent creates child agents with subAgent(), and each child has its own SQLite and execution context so data doesn't automatically mix. TypeScript type checking also catches incorrect RPC usage at compile time.

Conversation and memory are handled by the Session API. Messages are stored not as a linear list but as a tree with parent_id, enabling forking, non-destructive compaction, and FTS5-based full-text search. This design is meant to avoid losing context during long-running work.

The limitations of tool calling are addressed head-on as well. Cloudflare emphasizes a codemode approach that has the model write code instead of sequentially calling multiple tools. For example, a single program that finds and reads files to collect TODOs runs in a sandbox, and the Cloudflare API MCP server exposes only two tools, processing it in about 1,000 tokens — a claimed 99.9% token reduction compared to traditional per-endpoint tool calling.

Code execution happens in Dynamic Workers. These are V8 isolates created at runtime, described as about 100x faster than containers and up to 100x more memory-efficient. They start in a near-permissionless state with no network access by default, adopting a capability model where developers explicitly grant only the permissions needed via bindings.

All of these elements come together in the execution ladder:

• Tier 0: Workspace, a persistent virtual filesystem on SQLite and R2
• Tier 1: JS execution in a Dynamic Worker with no network
• Tier 2: An execution environment with npm added
• Tier 3: Browser automation
• Tier 4: Isolated sandbox

Finally, it also presents self-authored extensions — a direction where agents write their own tools at runtime, extending their capabilities as needed on the fly. In the end, Project Think proposes a foundation for treating AI agents not as applications, but as operable infrastructure.