What happened
On 25 June 2026, OpenAI published a piece on its site titled "How agents are transforming work." The article makes a single argument: AI agents, and in particular OpenAI's own coding agent Codex, are being handed bigger and more involved jobs than before, and by a wider mix of people inside companies rather than only software developers.
An agent here means a language model wired into a loop, so it can take several steps on its own toward a goal, calling tools and checking its own work along the way, instead of answering one question and stopping. OpenAI's claim is that, over the past year, people have started trusting this kind of tool with tasks that used to take a person half an hour or more.
To support that, OpenAI shares a set of usage figures drawn from its own product. By its account, active Codex users grew several times over in the first half of 2026, a larger share of requests was aimed at longer tasks, and use spread into non-technical teams such as Legal, Finance, and Recruiting. One detail is worth holding onto: every one of those numbers comes from OpenAI itself, measured on its own service. Same-day coverage of the piece, including from Axios and The Next Web, made the same point, noting that none of the figures have been checked by an outside party and that OpenAI both builds and sells the tool it is measuring.
So the honest summary is narrower than the headline. A company that sells AI agents published its own data arguing that AI agents are catching on. That is interesting and plausible, but it is a vendor's report about its own product, not an independent study of how work is changing.
Why it matters
This piece is a good moment to get clear on what an agent is, because the word is doing a lot of work in AI marketing right now. Picture an ordinary chatbot built on a language model: you ask, it answers, and the exchange ends. An agent is the same kind of model placed inside a loop. You give it a goal, and it can take several steps toward that goal on its own, run a tool, look at the result, and decide what to do next, before it comes back to you. That is the whole difference, and it is less dramatic than "autonomous worker" makes it sound.
What OpenAI is reporting, if the trend is real, is that people are comfortable handing these loops longer jobs than they used to. A year ago you might have asked an agent for a quick snippet. Now, per OpenAI's data, more requests look like work that would take a person half an hour or more. For a small team, that is the part worth noticing, not the exact percentages. It points to a change in how the tools get used: less "answer this one thing," and more "take this small, well-defined task off my plate and show me what you did."
The caution sits right next to it. An agent that takes several steps on its own is also an agent that can be confidently wrong for several steps, so the longer the leash, the more the review at the end matters. And because the adoption figures here are OpenAI's own, measured inside the company that profits from them, they are a reason to try the tools yourself rather than a reason to assume everyone else already has.
What to do next
- Treat the report as a vendor's case, not a verdict. OpenAI is describing its own product with its own numbers, and outlets like Axios and The Next Web flagged that the data is self-reported. Read it for the idea, not the statistics.
- Try an agent on one small, well-scoped task. Pick something you can check quickly, like drafting a short script or tidying up a file, and judge it on that result. An agent should earn trust on small jobs before you hand it larger ones.
- Keep a human review step at the end. Because an agent acts over several steps, a mistake made early can carry all the way through, so read what it did before you ship or send anything.
- You do not need to be a developer to start. The clearest claim in the piece is that non-technical teams are using these tools too, so if a repetitive workflow eats your week, it is reasonable to test whether an agent can take a piece of it.
This briefing summarizes a 25 June 2026 article that OpenAI published about its own agent products, along with same-day coverage from Axios and The Next Web, and links to those sources rather than reporting anything new. The usage figures it describes are OpenAI's own self-reported data and have not been independently verified.