Codex running for hours turns chat into delegated work

Ars Technica describes the new Codex as OpenAI’s rebranded tool for independent workflows that can run for hours if needed. The article is listed in metadata as 592 words and the available description frames the shift around work done for the user and with the user, not another autocomplete box.

Codex is moving from chat replies to long tasks

The available description suggests OpenAI wants Codex to handle more self contained work. The important word is hours. A short chat answer can be checked quickly. An agent that runs for a long time makes decisions, gathers context and can stack errors along the way.

That makes the rebrand more than cosmetic. OpenAI is saying the product should be a place where a team delegates work and returns at checkpoints, not just a place where a developer types a prompt.

Developer workflow breaks at the approval layer

For engineering teams, the impact is work control. If an agent prepares a change, tests, explanation and next step, a human is no longer approving one line at a time. They are approving the route the agent took.

That changes senior work. Less time may go into mechanical writing and more into task design, boundaries, testability and review. In a small team that sounds useful. In a larger company it immediately hits audit, permissions, repository access and rules for who may touch production code.

An agent that runs for hours needs brakes, not applause

The biggest risk is not that Codex writes one bad function. The worse failure is that it moves in the wrong direction long enough for the result to look polished and expensive. Longer runs require more checkpointing.

Marketing can show smooth delegation. Production will be duller: tests, sandboxes, limited permissions, logs, reproducible steps and a clear way to stop the agent. Without that, hours of autonomy become an expensive generator of review debt.

The useful metric is human interventions before merge

The metric to watch is simple: how many times a human must step into the agent’s work before the result reaches a merge request. If interventions fall while test quality holds, Codex is approaching useful delegation.

If teams merely move time from writing code to repairing agent decisions, this is a shiny interface over old labor. The early winners will be the teams with disciplined review, not the ones with the prettiest demo.

Lilith's verdict

A long running Codex is not a magical coworker. It is a junior with a badge to the repository, and the senior still needs to know when to stop him at the turnstile.