Can Codex Cloud Pass Your Worker Acceptance Test?

Run the same four permission checks and timed fixture in Codex Cloud and on your hosted Codex worker. Use the decision table to identify which environment meets your requirements.

Check the setup and task phases independently

Use a disposable repository with a tiny test that emits a known checksum. Put a harmless fixture file outside the approved checkout and define one permitted write inside it. Record four results for each environment: checkout write succeeds, outside write is blocked, setup can use an approved package fixture, and task execution blocks access to a separate team-controlled fixture. The last two are example policy requirements. Choose the required outcomes before the pilot and use identical requirements on both candidates. Where a fixture is supposed to be blocked, first prove that it is reachable from an authorized test context. A DNS failure or missing directory is an inconclusive result. Codex Cloud gives each task a workspace, so start each repeat as a new task and check whether the published setup produces the same checksum without relying on files from the prior run.

Codex Cloud | ChatGPT Learn

Time the path to a reviewable result

Start the clock when you submit the job and stop when a reviewer can retrieve the expected checksum and test output. Include queueing, setup and approval waits. Use several matched runs at the same time of day, then compare the slowest observed completion with your deadline. Keep approval events beside the timing record. A prompt that nobody can answer during an unattended run counts as unresolved, even if an operator can approve it later. Cursor documents a useful distinction for teams comparing architectures: its self-hosted machines execute tool calls locally while the agent loop remains in its cloud. For any hosted Codex worker, draw the equivalent map for your own setup and verify the destinations of the checkout, execution, and model requests separately. A pass on file permissions alone gives you evidence about just one of those boundaries.

Choose where Cloud Agents run | Cursor Docs

Choose what must survive the task

Codex Cloud gives each task its own workspace and can keep working after your computer sleeps. Start there if the job can begin from an approved repository and a reusable environment setup. Run the same task twice and check whether each run produces the expected patch and test results from the declared inputs. This establishes whether the job needs a long-lived machine at all.

Evaluate a hosted Codex worker when the next job needs a pinned system dependency, an existing build cache, or a file retained on the same machine. Define which items may persist and which should be reset. In a disposable checkout, ask for one expected file change and one attempted write to an approved scratch path outside the intended command boundary. Compare both results with the policy you meant to enforce.

Codex Cloud | ChatGPT Learn

Sources