OpenAI's Codex still asks a person to check the code it writes

A cloud agent announced 16 May 2025

On 16 May 2025, OpenAI introduced Codex as 'a cloud-based software engineering agent that can work on many tasks in parallel', built on a model it calls codex-1. The post states each assigned task 'runs in its own cloud sandbox environment, preloaded with your repository', where the agent can read and edit files and run commands such as test harnesses and linters. At launch, the same post says, 'internet access is disabled' during a task, limiting the agent to the repository and dependencies a user pre-installs — a scope OpenAI itself flags as time-bound, adding that the description 'describes the launch configuration' and pointing readers to separate, current documentation for how networking works since. OpenAI's own site now marks this announcement 'outdated', a useful reminder that a vendor's launch post is a record of one date, not a running description of the product.

Parallel tasks change what a person has to check

Because Codex can be assigned several tasks at once, each producing its own committed changes, the announcement is direct about where responsibility still sits: 'it still remains essential for users to manually review and validate all agent-generated code before integration and execution'. A person can 'review the results, request further revisions, open a GitHub pull request, or directly integrate the changes' once a task finishes, but nothing in the announcement describes an automatic check that substitutes for that review. Running five tasks in parallel does not remove five reviews; it moves them later and bunches them together, a workload question a team adopting parallel agents has to plan for rather than assume away.

What the benchmark and case studies do not establish

The announcement cites 'internal benchmarks' and a curated 'internal SWE task benchmark' built by OpenAI itself, plus a SWE-Bench Verified result from which the post says '23 SWE-Bench Verified samples that were not runnable on our internal infrastructure were excluded' — a methodology choice made and reported by the vendor, not verified by an outside party. The named early users quoted in the post, including Cisco and Superhuman, were selected by OpenAI to describe their own experience, which is evidence of adoption, not a controlled comparison. Separately, Codex CLI, a related but distinct terminal tool, has its own current documentation, which describes a '/permissions' command letting a developer 'choose when Codex can edit files or run commands without asking', and a review command that 'reports prioritized findings without modifying your working tree, so you can address risks before you commit or open a pull request' — a check placed explicitly before commit, in contrast to the cloud agent's after-the-fact review.

• When several agent tasks run at once, who is scheduled to review each result, and by when?
• Does a cited benchmark figure exclude cases the vendor's own infrastructure could not run?
• Is a customer story a controlled comparison, or a vendor-selected account of one team's experience?

Sandboxing addresses where an agent can act, not whether its output is correct. Parallel execution changes the shape of a team's review queue more than it changes the need for one.