Agent changes need isolated runtime previews

Summary

Cloudflare Worker Previews give a Git branch a running Worker with its own URL, configuration, Durable Object and Container state, and observability. This brings runtime checks into the branch review loop, including for agent-authored changes. It does not isolate the full application path: service bindings still call bound Workers in production, previews cannot consume Queue messages, and Workflow executions need separate configuration for isolation.

Key Updates

• A preview can be created with npx wrangler preview; a Git-connected Worker can create one on push through Workers Builds. Each branch keeps a stable preview URL that updates with new pushes.
• Preview settings can be defined in a previews block and overridden for one preview, including a test database or API key. Cloudflare creates separate Durable Object namespaces and Container applications for each preview.
• Logs, errors, metrics and traces are scoped to the preview, so reviewers can inspect runtime behavior before merge.
• Cross-service limits remain: service bindings target production Workers; previews can send Queue messages but cannot consume them; Workflow isolation needs separate configuration.

Why It Matters

As agents contribute more changes, teams need to run and inspect each change without relying on a shared staging site. Worker Previews provide that environment for a branch's Worker and its supported stateful resources. The release also shows where a preview stops being an isolated test: a request that crosses into a bound Worker or an asynchronous path may reach production or miss the behavior under test. A successful preview is evidence for the paths exercised, not proof that the entire system was isolated.

Builder Takeaway

Map the bindings and asynchronous steps in the change before testing. Use preview-specific configuration and test resources where available, and check whether service-bound calls still reach production. Exercise the relevant HTTP path, inspect preview traces and errors, and test Queue consumers and Workflows through a separate controlled setup. Record which paths the preview actually covered.