In July 2026, OpenAI made its Codex desktop app the new ChatGPT desktop app, with chat built into it. A tool made for programmers is now the base of the product that hundreds of millions of people already use. The idea behind that move is more useful to an office worker than to a programmer: an agent gets good at a kind of work when two things are true. It can reach the files and tools where the work lives, and the result can be checked fast.
Code came first because code checks itself
Programming is the easiest work to hand to an agent, because it is a closed loop. The agent changes the code, runs it, and sees within seconds whether the change worked. If it did not, the agent backs up and tries another way, and nobody has to look over its shoulder.
Most office work has no such loop yet, but some of it is close. Take a bookkeeper who spends two days a month matching 400 supplier invoices against bank lines. An agent with access to the spreadsheet and the bank export can do the matching, and the bookkeeper checks it by looking at whether the totals agree and reading the short list of items that did not match. That is perhaps a day and a half returned each month, and the checking stays with the person. One in five Codex users is already a knowledge worker rather than a coder, so this is not a distant case.
The first week feels worse, and then it flips
The Codex team made a choice that looked wrong. Their editor extension was the most-used way to run Codex, and usage had grown more than 10-fold. Instead of building on it, they made a desktop app with no code editor at all. If a user wanted to change a line, they had to ask the agent.
At first the team's own members avoided it. One engineer wrote that prompting the agent just to replace a sentence felt cumbersome. A few weeks later, once the models improved, the same people were leaving comments on the agent's work and no longer missed typing the change themselves. Taking away the manual option had forced the product to make the agent good at everything the user might need.
Anyone starting to delegate at work will go through the same week. Doing the small fix by hand feels faster, and it keeps the agent from ever learning your files and habits. A better test is to give it one whole task whose result you can check in a minute, and to fix mistakes by telling it rather than by editing.
The cloud comes back, in a different order
OpenAI plans to move Codex back to the cloud, so that agents keep working after you close the laptop. That is the idea that failed in 2025. This time it comes after millions of people have learned to trust an agent on their own computer, with their own files, and have seen how its work gets checked. It fits the always-on agents OpenAI has already started putting in ChatGPT, and the work tools we covered in August.
The order is the part to copy. Trust built on a machine you control, with work you can check, is what makes it reasonable to let an agent run unattended later.