Cloudflare wants coding agents that survive the crash

A new experimental harness runs the Codex agent loop on Cloudflare - and the caveats are the interesting part.

Today Cloudflare published an experimental example that runs OpenAI's Codex agent loop inside a Cloudflare Durable Object. Conversations, files, and completed model and tool results persist across restarts. After an interruption, the agent resumes from its saved state.

That matters more than it sounds. Coding agents fail in boring ways: timeouts, restarts, dropped connections. A demo that runs for five minutes is easy. An agent that survives a crash and keeps its state is the actual product problem.

How it works

The example adapts Codex's Rust turn-processing code to WebAssembly and builds on four pieces of Cloudflare's Agents SDK:

• Tasks journals model calls and tool execution. Completed steps reuse their recorded results after a restart.
• Sessions stores conversation history and compacts the model context.
• Streams stores ordered events so a client can replay them after reconnecting.
• Workspace lets the model read and write persistent files.

The caveats (read this part twice)

• It is experimental. The harness is example code built on internal SDK APIs, not a supported product. A starting point, not a foundation.
• Recovery is not exactly-once. Completed steps reuse stored results, but an interrupted step - including a model call - can run again. A tool with external side effects, like charging a card or sending a message, needs its own retry and deduplication strategy.
• It is not free to run. Model calls are billed through Cloudflare's AI Gateway, either with Unified Billing or your own OpenAI API key - even during local development.
• The example routes are unauthenticated. Cloudflare's own docs say to add authentication before exposing user sessions.

Our read

Editorial interpretation, not Cloudflare's claims. Durable execution used to be a backend niche. We read today's release as a signal that in the agent era it becomes a selling point: whoever makes long-running agents reliable - restarts, retries, replay - owns infrastructure every agent product needs. Cloudflare appears to be betting that this layer belongs at the edge.

If that read is right, today's release is less a product than a signal: a working coding agent assembled from primitives Cloudflare now ships. In our view, the moat is not the agent loop. It is the durability layer underneath it.