AI Briefing
KO

Project Think: Building the Next-Generation AI Agents on Cloudflare

·2026.04.16 10:52

Key point

Cloudflare has unveiled a next-generation Agents SDK equipped with durability, sandboxing, and sub-agents.

Details

Cloudflare has extended its existing Agents SDK through Project Think. The goal is to provide fundamental primitives for long-running agents, allowing developers to build safer, more scalable agents with less configuration.

The core idea is treating agents as infrastructure-level units of execution. Each agent operates as an actor model based on Durable Objects, with zero compute cost while dormant, and automatically wakes up when a message or alarm arrives. Cloudflare explains that by providing per-agent SQLite state, name-based routing, automatic restarts, and state preservation, this is far more economical at large-scale concurrent sessions than an always-on container architecture.

The key primitives are as follows.

  • Fibers: Durably execute long-running tasks via runFiber(), with support for checkpointing and recovery
  • Facets: Co-locate sub-agents isolated from the parent to separate tasks
  • Session API: Stores conversations in a tree structure, supporting forking, summary compression, and FTS5 search
  • Dynamic Workers: A sandboxed JS execution environment with no network access by default
  • Execution Ladder: Progressive scaling from Workspace → Dynamic Worker → npm bundling → headless browser → Cloudflare Sandbox

Cloudflare specifically believes that having an LLM write a single program to handle a system is more efficient than the existing repeated tool-calling loop. To this end, it provides standalone packages such as @cloudflare/codemode, @cloudflare/worker-bundler, and @cloudflare/shell, and proposes a structure that only escalates to a stronger execution environment when needed.

Another axis is self-authored extensions. Agents can directly create and deploy new tools during execution, and these extensions are sandboxed, auditable, and revocable within a Dynamic Worker. As a result, agents evolve into something closer to a serverless execution platform equipped with persistence, isolation, security, and scalability, rather than a simple chatbot.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.