Hermes picks up the phone
Every agent you have used waits for you. You open the app, you type, it answers. The initiative is always yours.
We wanted the opposite: an agent that decides it should reach you, and calls. Not a notification. A phone call. It rings, you answer, and you are talking to it.
Hermes can do that now.
What it does
- It calls you. On its own schedule, or when something matters, Hermes places a real call and starts talking the moment you pick up.
- It briefs you. "Here is what is ahead today" — hands-free, while you are making coffee, no screen involved.
- You talk back. Full duplex. Interrupt it mid-sentence, ask a follow-up. It listens and responds in real time, not walkie-talkie style.
- It knows its limits. On the call Hermes is the voice, but it checks with its own brain before stating facts about your world, and it never takes an action without confirming first. It can chat freely. It cannot freelance.
- Only you can reach it. The line is locked to a single identity, end to end. A stranger who finds it hits a wall.
Why voice is different
Text agents are reactive by design. They sit in a thread until summoned. Voice flips the relationship: the agent can be proactive. It reaches out at the right moment, in the one channel you cannot ignore and do not need a screen for. A morning brief you listen to beats a morning brief you have to open.
And a call is a conversation, not a transcript. You interrupt, you clarify, you decide, at the speed of talking.
How we built it
Three pieces, and the interesting part is how little glue they need.
- A realtime, speech-to-speech voice model (Grok Voice) is the ears and the mouth. It hears you, reasons, and speaks — sub-second, no awkward pauses, and it handles being interrupted.
- The agent's own brain stays the authority. The voice drives the conversation, but for anything factual it asks the brain, and for anything consequential it waits for the brain's go-ahead. The voice is the front office. The brain is the chief.
- A thin bridge carries audio between the phone and the model. The voice model speaks the same audio a phone speaks, so the bridge mostly passes bytes straight through.
No shared runtime, no model retraining, no bespoke telephony stack to babysit.
The pattern
The lesson generalizes past phone calls. Give a capable voice model the audio and let a separate, trusted brain hold the authority — the voice for fluency and turn-taking, the brain for truth and permission. You get an agent that is pleasant to talk to and that you can trust not to act on its own.
An agent that only answers is a tool. An agent that calls you when it matters — and knows what it is allowed to do once you pick up — starts to feel like staff.
More on the Hermes control plane: Steer your autonomous engineer from chat.