AI Agents Can Now Ship Code to Production on Their Own

Cloudflare has introduced temporary accounts that let an AI agent deploy code to its network without ever signing up, logging in, or clearing a two-factor prompt. The mechanism is deliberately blunt: an agent runs wrangler deploy --temporary, and Cloudflare instantly provisions a throwaway account, hands back an API token and a claim URL, and pushes the code live. The deployment stays up for 60 minutes, during which a human can claim the account to make it permanent — or do nothing and let it auto-expire and delete itself.

The reasoning is a small but telling observation: nearly every step of cloud onboarding — OAuth flows, dashboards, API-token management, MFA — was designed for a human sitting at a keyboard. An autonomous agent or a background process has no hands to click 'approve' and no phone to read a code from. That mismatch quietly blocks the exact trial-and-error loop coding agents are good at. When a first deploy fails on authentication, Wrangler now nudges the agent toward the --temporary flag instead of dead-ending.

Agents get a sandbox that's also real

What makes this more than a gimmick is that the temporary deployment is a real one. Within the 60-minute window the agent can iterate — redeploy code changes again and again — and it isn't limited to toy pages: it works with Cloudflare Workers, full websites, APIs, and resource bindings like databases. So an agent can stand up a working app, wire it to a database, test it, and only then hand a human the claim link to keep it. The clock and the auto-delete are the safety rail: nothing lingers unless a person decides it should.

It ships in the latest Wrangler CLI release, so it's available to anyone whose agent already drives Cloudflare's tooling.

Why it matters for you

If you've tried to have an AI coding assistant actually deploy something, you've probably hit the wall where the agent writes working code and then stalls at a login screen it can't pass. This removes that wall for one common target. More broadly, it's an early example of infrastructure being reshaped for non-human users — a hint of where 'agent-first' developer tools are heading, where the signup-and-auth ritual is optional rather than mandatory.

My take: the 60-minute auto-expiry is the part I'd actually trust. Letting an agent create accounts sounds alarming until you realize nothing it makes survives unless a human claims it — that's a sane default, and it's the design detail I'd want copied if other platforms follow.