Talk to the Agent While It Works

On November 7, 2026, Cursor and Google each announced two-way voice for coding agents. The New Stack covered both the same day the news landed for readers: Cursor's Voice Chat with Agents runs on Grok Voice, and Google's Antigravity voice mode uses Gemini Live in the desktop app and CLI.

The pitch is similar in both products. Working with an agent has had a typed rhythm: prompt, wait, read, type again. Longer jobs break that rhythm. A command needs approval. A requirement was missing. You notice a better approach halfway through. Each of those moments pulls you back to the keyboard.

Voice is one answer. Keep a live call open. Hear progress. Steer without leaving what you were looking at.

Who is talking, and who is not#

OpenAI already had two-way voice in Codex, including in the ChatGPT desktop app and the Codex CLI. Cursor added spoken progress updates earlier and is now pushing a fuller call experience across desktop, web, and iOS. Google is consolidating coding work into Antigravity and treating voice as a live session, not only dictation.

Anthropic's Claude app has two-way voice, but Claude Code remains dictation-only. Speech becomes prompt text. The agent still replies in writing.

That split is useful. Dictation helps you describe a bug or a feature. Two-way voice helps you stay in the loop while the agent is already working. Those are different interface problems.

Voice is a presence problem#

The deeper issue is not microphones. It is how an agent shows state when the work takes minutes, not seconds.

If the only signal is a spinner, people interrupt out of anxiety. If every step is dumped into a wall of logs, people drown. Good agent interfaces answer three questions quickly:

1. Who is this?
2. What are they doing?
3. How much do I need to know right now?

Voice can carry the second question without forcing you to watch the transcript. A short spoken update is often enough. Hovering a status, opening a preview, or taking over a computer view can carry the third. The interface should default to reassurance, then offer depth on demand.

That is also why chat is not always the right surface. VS Code 1.144 added canvases for richer shared artifacts for the same reason. Some work wants a board, a diff, a form, or a live preview. Prose is one shape of answer. It should not be the only shape.

What the model side enables#

The speech models are getting good enough that the interface question is no longer theoretical. Grok Voice 3.0, launched October 14, 2026, ranked first on Artificial Analysis's Speech-to-Speech Index with a score of 86.9. It posted a 94.8% task success rate and 97% on Full Duplex Bench, which measures turn-taking, pauses, and interruptions. Pricing stayed at $0.08 per minute.

When latency and turn-taking are good enough, product teams have to decide what the conversation is for. Is voice a faster prompt box, or a pairing channel that stays open while tools run?

Design choices that hold up#

I would keep a few rules close while this pattern spreads.

Prefer short spoken status over narrating every tool call. People want to know the agent is alive and on track. They do not need a radio play of npm test.

Make interruption cheap and safe. Mid-task steering only works if the agent can absorb a correction without throwing away useful progress.

Keep a written record. Voice is for the moment. The transcript, the diff, and the decision log are for later.

Separate dictation from dialogue. One turns speech into a prompt. The other keeps a channel open during execution. Ship the one you mean.

Talking to an agent while it works will feel normal soon. The teams that get it right will treat voice as part of presence and control, not as a novelty input method.