Hands Off the Keyboard, Drive Your Agents by Voice — OpenAI's "Full-Duplex" Voice Control on Codex/Work, and the Slip-Ups Its Speed Invites

Between July 23 and 24, 2026, OpenAI rolled out "ChatGPT Voice," a voice-control feature for the ChatGPT desktop app (macOS / Windows). The goal isn't casual chat. It's to let developers instruct multiple ChatGPT Work or Codex agents and have them write code using nothing but their voice. Here we break down how this feature changes the flow of development work, and the pitfalls that come with pouring in instructions by voice.

The voice is decoupled from the "executor"

At the core is "GPT-Live," the voice model that debuted on July 8. It uses a full-duplex approach that lets it listen and speak at the same time, doing away with the old turn-taking of "speak → stop → respond." It can slip in an acknowledging "Got it" while taking in your instructions without cutting you off.

The key point is that this voice layer is separate from the execution engine that actually does the work. GPT-Live handles only the conversation, passing heavy reasoning and code generation to the models behind it and to the Work / Codex agents. While you use your voice to "start work, reprioritize, stop, or reassign," the agents keep running in the background and report stalls or completions by voice or on screen.

In OpenAI's demo, a developer used a single voice command to spin up a new thread, create a pull request, and pin down the root cause of a bug. Hands stayed off the keyboard while instructions flowed by voice alone.

What changed on the Codex app side

The Codex app update (26.715) shipped alongside it broadened the handling of local projects on top of voice. There are two changes:

• Added voice control: Using GPT-Live, you can drive work by voice across Chat / Work / Codex in the desktop app.
• Multi-folder projects: A single local project can now include multiple related folders. You designate one main folder, which becomes the starting point for new chats, Git operations, and auto-detection of AGENTS.md / skills / config.toml. Sub-folders remain available as targets for file search, reading, and writing.

On macOS, enabling "screen context (Appshots)" lets you share the contents of an open window and fold it into the context. In other words, you can layer the state of the screen you're looking at onto your voice instructions.

Availability and requirements

The coverage and operating conditions are as follows. Before adopting it, check whether your plan and environment qualify.

Voice works together with Computer Use, pre-authorized tools, and local files. It doesn't sprout new, unlimited permissions; the closer read is that "voice" has been added as an entry point on top of your existing authorizations.

From "queue up and wait" to "drive while you talk"

Until now, running agents basically meant typing out instructions, queuing up several tasks, and waiting for results. This voice control brings in a new move: "swapping priorities out loud while things are still running." Being able to interrupt multiple agents without stopping your hands looks like the practical difference.

On the business side, the effort spent on routine handoffs like reviews, PR creation, and debugging is likely to drop. Since you can step away from your desk and check on things by voice through Remote on iOS, how you use waiting time changes too. As the starting point of work shifts partly from the editor to "voice," it pushes up the number of agents a single developer can keep an eye on at once.

The very ease of handing off is what breeds slip-ups

Behind the convenience, risks unique to voice have been flagged. Without overhyping them, here are the points worth keeping in mind when adopting it.

• Transcription drift: When speech overlaps or noise creeps in, the audio and the transcription can diverge. It's safer to leave agreed-upon terms and conditions in text on the task itself, rather than relying on the voice record alone.
• Ambiguous instructions spread fast: Voice lowers the friction of handoffs, but it also raises the speed at which a vague spoken instruction fans out across multiple tasks at once. The structure can make rework prone to chaining.
• Data retention: Voice clips are said to be retained for 30 days, so how deletion and archiving are handled needs to be confirmed in your operating rules.
• Screen-sharing context: When you use screen context, the shared content mixes into your instructions. The windows you open and anything displayed from external sources warrant the same care as before.

The fact that permissions stay confined to the scope of your existing context acts as a brake, but "being able to hand off fast" itself makes it easy to skip checks. Lock the conditions down in text before you drive by voice — that one bit of extra effort looks like it pays off.

In summary

This update extends the control surface of coding agents from "typing" to "talking." The experience of running multiple agents in parallel with full-duplex voice does genuinely speed up the flow of work. At the same time, it brings along new slip-ups: the ambiguity of spoken instructions and the drift in the record. The realistic way to start is to try voice on limited tasks first, and adopt a practice of committing the important conditions to text.

References: OpenAI official announcement (X) / TechCrunch / 9to5Mac / XenoSpectrum / Codex changelog (ChatGPT Learn)