
Cloudflare introduced the open beta of cf on September 28, 2026, with an agent-oriented command interface and JSON output. A useful first trial for a website team is a bounded resource inventory, before giving an agent permission to change anything.
1. Choose an inventory you can check
Pick one disposable zone or test environment and write down what the report should contain: resource types, names and stable identifiers. Keep purchases, deployments and DNS edits outside this exercise. The procedure below is our suggested acceptance test, not a claim that cf enforces these restrictions by default.
Create a technical identity with only the read permissions needed for this scope. Apply resource restrictions where supported. Supply credentials through the tool’s supported authentication mechanism; do not paste them into prompts or screenshots. Confirm the selected account before starting.
Record the installed version as well. Beta interfaces can change, and generated commands should be checked against the help for the version actually installed, rather than accepted because they look plausible.
2. Specify how incomplete results must appear
Ask the agent to propose its steps first. Permit only the reads needed for the approved inventory. This example is a task contract, not executable cf configuration:
Target: one test environmentReport: type, name, stable IDDenied read: mark inventory incompletePagination: report completion stateWrite operation: stopRequire a clear distinction between an empty list and a failed request. Missing permission must not become “no resources found.” If an endpoint returns several pages, the agent should explain whether it retrieved all of them.
Structured output helps you process results, but it does not prove that the correct account was queried. Retain request status and useful identifiers alongside the inventory. Remove sensitive fields before passing responses to an external model, and avoid retaining entire responses simply because they are JSON.
Set a reasonable time limit for this small trial. If it expires, an incomplete report with the last successful step is more useful than a confident summary built from partial results.
3. Accept the pilot using independent evidence
In the dashboard, select a few known resources, including a recent addition. Match their identifiers to the report. Also check that a deliberately excluded test resource remains outside the identity’s reach.
Repeat a read with a required permission removed in the test environment. The agent should report the denial and stop instead of silently broadening its scope. Save the outcome, tool version and date as the baseline for later trials.
Approve this pilot only when account selection, pagination and permission failures are represented correctly. Introducing write access later deserves a separate acceptance test with a named reviewer and an explicit recovery procedure.
Sources: Cloudflare’s cf announcement.