Table of Contents
ToggleWhen 100 million knowledge workers each need a resident AI assistant, traditional container architecture has a boring but fatal problem: it's burning money no matter whether there's anything running in the container or not.
With 10,000 agents, each only active 1% of the time, container deployment would require maintaining 10,000 constant instances. To save resources, Cloudflare released Project Think this month, positioned as the next-generation version of the Agents SDK, and open-sourced it on GitHub .
The company claims that its Agents SDK is already driving "thousands" of agents in production environments. Project Think builds on this foundation by introducing six new underlying primitives, addressing not the cognitive abilities of AI, but rather the three infrastructure issues of "sustainability, security, and cost."
What is Project Think? A breakdown of its six new Primitives
I. Durable Execution via Fibers
Traditional stateless functions crash halfway through execution, losing all previous progress. Fibers are designed so that `runFiber()` writes the state to an SQLite instance before execution, and the agent can store the current state at any time via `stash()`. In case of a crash, execution can resume from `onFiberRecovered`. This is a fundamental condition for enabling long-running tasks (such as research agents running for hours) to survive in real-world environments.
II. Sub-agents
Child proxies exist as co-located Durable Objects, each with its own independent SQLite database and execution environment, maintaining co-location with the parent proxy through the Facets mechanism. TypeScript can detect misuse at compile time, rather than throwing errors at runtime.
III. Persistent Sessions
Each message has a parent_id, forming a tree structure. It supports fork (branching exploration without disrupting the original dialogue path) and non-destructive compaction (summarizing old messages instead of deleting them), and has built-in FTS5 full-text search. This solves the problem of "AI assistants losing their memory after multiple conversations".
IV. Sandboxed Code Execution
Using Dynamic Workers (based on V8 isolate), startup time is measured in milliseconds, making it 100 times faster and 100 times more memory efficient than traditional containers. The capability model uses explicit licensing: the default is globalOutbound: null, meaning developers must explicitly grant access permissions to each resource, rather than defaulting to open access.
V. The Execution Ladder
The five tiers are arranged from lightest to heaviest: Tier 0 is SQLite + R2 virtual file system (pure data storage and retrieval), Tier 1 is sandboxed JS execution, Tier 2 adds npm parsing and esbuild packaging, Tier 3 is a headless browser (Browser Run), and Tier 4 is a complete Cloudflare Sandbox (supporting git clone, npm test, and cargo build). The agent chooses which tier to enter based on the complexity of the task, rather than defaulting to the heaviest one.
VI. Self-authored Extensions
Agents can write their own TypeScript extension tools at runtime, which are then packaged into the Dynamic Worker by the ExtensionManager. Once a new tool is created, it exists permanently—no fine-tuning or RLHF is required; the tool knowledge resides in the infrastructure, not the model.
Think Base Class: Unified Lifecycle
Project Think provides a Think base class that encapsulates the complete lifecycle of a chat agent: loop, persistence, streaming, tool execution, and recovery after interruption. Subclass inheritance requires only a minimum of 5 lines of code.
Built-in features include context-block memory, tree-structured sessions, FTS5 full-text search, and non-destructive compaction.
The design intent is clear: to remove the cost of reinventing the wheel from each developer and let the infrastructure bear the complexity of state management.
Cloudflare's Three Waves Theory
Cloudflare proposed a three-wave evolution of AI Agents in its documentation: the first wave is stateless chatbots, the second wave is code agents that have tools but are stuck on local laptops (Claude Code, Codex, etc.), and the third wave is "agent as infrastructure": persistent, serverless, zero idle cost, and structurally secure isolation.
Project Think is the third wave of infrastructure bets. Cloudflare's argument is: to move tools like Claude Code from laptops to the cloud, to enable them to recover after crashes, and to make the cost of 100,000 agents equal to the actual usage rather than the maximum number of parallelisms.
It's worth noting that Cloudflare's own AI Gateway and Workers AI are also part of this ecosystem. The success of Project Think also translates into traffic and revenue for the Cloudflare platform. This three-wave narrative is essentially a market positioning manual.
