Cloudflare Ships Five Agent Products in Three Days — Project Think, Mesh, and the Bid to Own Agent Infrastructure
Between April 15 and 17, Cloudflare shipped five distinct agent-related products under the banner of “Agents Week.” Taken individually, each is a product launch. Taken together, they represent a deliberate attempt to become the default infrastructure layer for production AI agents — the same position AWS filled for web applications in 2010.
Project Think: The Agents SDK Grows Up
The headline product is Project Think, a next-generation Agents SDK built on Cloudflare Workers and Durable Objects. It ships four new primitives:
- Durable execution with fibers — crash recovery, automatic checkpointing, keepalive. Agents that were previously killed by a dropped connection or cold-start restart from where they left off.
- Sub-agents — typed RPC between isolated child agents, enabling multi-agent workflows without shared state risk.
- Sandboxed code execution — LLM-generated code runs inside Dynamic Workers, separated from host infrastructure. Packages resolve at runtime.
- Persistent sessions — every agent gets a durable virtual filesystem backed by SQLite and R2 that survives restarts.
The Think base class wires all four together. The pitch: an agent running on Cloudflare now has the same execution model as one running locally — file system, code execution, browser access, scheduled waking — but serverless, durable, and globally distributed.
Cloudflare published the context directly in the blog post: tools like OpenClaw, Claude Code, and Codex proved that giving a model a real environment to act in changes what it can accomplish. Project Think extends that pattern to hosted infrastructure.
Cloudflare Mesh: Private Networking for Agent Fleets
Alongside Project Think, Cloudflare launched Mesh — described as the first private networking solution built specifically for AI agents at scale. The problem it targets is one every team shipping production agents hits: agents need access to private databases, internal APIs, and staging environments, but granting that access has previously required exposing infrastructure to the open internet.
Mesh creates a private fabric that routes agent traffic through Cloudflare’s global network without traversing the public internet. Each agent in a Mesh environment carries a distinct cryptographic identity. Security teams can write granular access policies: a coding agent can read a staging database; it cannot touch production financial records. The Mesh integrates directly with Workers VPC bindings, so agents running on Cloudflare Workers can access entire private networks through the same binding model used for storage and KV.
Agent Readiness Score and Redirects for AI Training
Two additional products target the web-owner side of the agent economy.
The Agent Readiness Score, published April 17, gives site owners a structured measure of how well their site supports AI agent access — structured data, canonical URLs, sitemap compliance, MCP endpoint presence. Cloudflare published Radar data showing significant variation across major sites; its own developer documentation was singled out as the highest-scoring property tested.
Redirects for AI Training addresses a gap that robots.txt was never designed to close. When a site redirects or deprecates a URL, crawlers often continue ingesting the old content — meaning AI training sets pick up stale or superseded versions. The new tool intercepts verified AI training crawlers at the edge and transparently redirects them to canonical pages with a single toggle, no origin changes required.
Agent Lee: 18,000 Daily Users, 1 Million Tool Calls
Rounding out the week was general availability of Agent Lee, Cloudflare’s in-dashboard AI assistant. During its active beta, Agent Lee served over 18,000 daily users and executed nearly 1 million tool calls — setting up routes, diagnosing Worker errors, applying Access policies. The system operates against live account data, not FAQ text.
Agent Lee is available in beta on free-tier Cloudflare accounts.
What This Adds Up To
The pattern across all five products is consistent: Cloudflare is positioning itself where agent traffic crosses infrastructure. Mesh handles private connectivity. Project Think handles execution and durability. The Readiness Score and Redirects tools handle the web-owner relationship with crawlers. Agent Lee handles the operator’s relationship with Cloudflare itself.
The existing agent platforms — Anthropic, OpenAI, Google — provide the models. The cloud providers — AWS, Azure, GCP — provide raw compute. Cloudflare is building for the connective tissue: identity, networking, execution, edge delivery, and crawler governance. If agents become the primary unit of software deployment, that connective tissue is where margin accrues.
Key Releases
| Product | Launch Date | Core Function |
|---|---|---|
| Project Think | April 15 | Agents SDK: durable execution, sub-agents, sandboxed code |
| Cloudflare Mesh | April 15 | Private networking fabric for agent identity and access |
| Agent Lee GA | April 15 | Dashboard AI: 18K daily users, ~1M tool calls in beta |
| Agent Readiness Score | April 17 | Per-site scoring for AI agent compatibility |
| Redirects for AI Training | April 17 | Edge-level canonical redirect for verified training crawlers |