Cloudflare has enabled AI agents to provision cloud infrastructure end-to-end — creating accounts, registering domains, initiating paid subscriptions, generating API tokens and deploying code — without repeated human intervention. The capability, currently in beta, requires a human to accept terms of service once, after which an agent can operate independently.
Co-designed with Stripe and reported by Computerworld, the announcement marks a material shift: AI is no longer just executing tasks inside existing workflows. It is now authorised to enter commercial agreements on your behalf.
What Cloudflare actually launched
The beta builds on a new protocol co-designed with Stripe acting as identity provider. Agents query a service catalogue, receive credentials, and obtain payment tokens to initiate subscriptions across providers including AgentMail, Supabase, Hugging Face and Twilio. Credentials are stored for ongoing agent use.
Cloudflare set a default spending cap of $100 per month per agent per provider, with budget alerts available for users who want to raise the limit. The company is also offering $100,000 in Cloudflare credits to startups that adopt the capability through Stripe Atlas, its company incorporation programme.
The governance risks businesses can’t ignore
Security practitioners raised concerns alongside the announcement: the same speed that makes autonomous provisioning useful also makes it dangerous. A misconfigured agent — or a malicious one — could spin up infrastructure at a rate no human oversight process could intercept in time.
The spend cap provides financial guardrails, but it does not limit the number of accounts, domains or API tokens an agent can create. For organisations already thinking about deploying AI agents, this introduces a new credential category — agent-held — that most existing identity governance frameworks were not built to handle.

What this means for Australian businesses
Finance and operations teams will need to answer questions that most policies don’t yet cover: who owns the spend an agent incurs, what approval process applies before an agent is granted provisioning authority, and how billing disputes across multiple providers get resolved.
Australia’s Privacy Act 1988 and ASD cloud security guidelines do not yet address autonomous agent provisioning specifically, but the accountability obligations they impose on data controllers apply regardless of whether a human or an agent initiated the activity. Businesses considering AI agent development should treat provisioning authority as a governance decision, not a technical one.
A shift that governance hasn’t caught up to
Agentic AI is being granted commercial authority faster than the frameworks to govern it are being written. Whether that represents acceleration or recklessness depends on the controls an organisation puts in place beforehand.
For Australian IT and finance leaders, the immediate priority is reviewing identity, credential and billing governance before autonomous agents reach production — not after the first unexpected invoice lands.
