OpenAI shows how to work with Codex by voice, from the cloud and through a personal agent called dot

A year ago, a coding agent was a window where you typed a request and waited. Now frontier labs are racing to offer a different mode: the agent works in the cloud, listens to your voice and lives between your sessions. On the first day of the AI Engineer Code Summit, OpenAI engineers Dominik Kundel and Charlie Guo showed how they work with Codex themselves. Richard Macmanus published a report from the workshop on Latent Space.

A cloud you can finally stop worrying about

Codex launched in May 2025 as a cloud agent for parallel tasks. But, as Kundel admitted, it took the models and the surrounding tooling time to turn all of this into a convenient experience. A noticeable shift came after DevDay in late September, when the second version of Codex Cloud was announced. The promise was simple: you can close your laptop, and the agents will keep working.

It works like this. You publish an environment once, meaning you pick the GitHub repositories you need access to, and then launch as many tasks as you like in isolated copies of that setup. An environment can be personal or shared across the whole company. For a team, it looks like a ready-made image of a workstation that doesn't need to be rebuilt for every task.

The most interesting part is moving work between local Codex and the cloud. The author of the report candidly admitted he still gets confused by this, and Kundel agreed: it should be simpler. His own trick is this: if a task started locally, he simply asks local Codex to push the branch and open a new cloud thread. In his words, the cloud becomes "just another executor," and local and cloud Codex are gradually merging. The difficulty, as he noted, is the difference in context between them.

Voice as a control panel

In the demo, Kundel edited the website of a fictional label, Blossom Music, by voice. Dictation is handy here when you don't yet fully understand what result you want and are iterating quickly.

Guo went further. In his observation, most people see voice as just another way to type on top of regular work. He uses voice as a control plane, that is, as a layer for managing everything: checking what is happening across threads, creating a new thread, pulling data from other apps on the computer. Not "one lane of work done by voice," but management of all tasks.

Kundel also uses voice this way: he talks to his dot on the way to work and kicks off PRs so he can review them later, at home or in the office. It's a powerful scenario, but it's important to remember verification here: the agent prepares, the human reviews.

dot: a personal agent between sessions

The workshop touched on personal agents like OpenAI's dot (Anthropic has an analogue called Claude Duet) only in passing. Guo said outright that for an experienced Codex user, dot feels unfamiliar: it's a new paradigm, and it is good at different things than the main Codex.

Nevertheless, he named three roles in which dot is useful:

• a safety net: the agent highlights what you may have missed;
• a "delegation sink": Guo often uses it in Slack;
• omnipresence: the goal is for dot to be everywhere you work.

Kundel starts many tasks with dot rather than with a prompt in Codex. The reason is that dot already has access to dev machines, the laptop and cloud environments, and therefore the context of what you are working on.

What to do with this

If you already use coding agents, there are several takeaways from the talk:

1. Set up a cloud environment once, especially if you work in a team. It removes repeated configuration.
2. Try voice not only for dictation but also for managing streams of tasks.
3. If you have a personal agent, give it the role of a reminder and a delegation inbox rather than a replacement for your main tool.
4. Follow Kundel's advice: lower the effort level when a task is simple, and ask the agents themselves to help with setup.

The limitations are worth stating too. This is a report from a workshop, not an independent test: there are no figures on speed or quality, and the transition between local and cloud work, according to the developers themselves, remains unintuitive. Kundel's final piece of advice is perhaps the most honest: don't get too attached to your current way of working with Codex, because everything is changing fast.