TL;DR: Agentic AI quietly changed what a keyboard is for. You no longer type the work; you describe the work, and descriptions are long, conversational, and constant. That makes voice the natural input for AI agents, and the hands-free loop is what makes several agents practical at once: say "type" plus your instructions and they get typed, say "send" and Enter is pressed, say "open Cursor" or "open Claude Code" and you are in the next session, all from a couple of feet away without touching a key. Infina is the Mac app built around exactly that loop, on-device by default, $99 once as of July 2026 with a 7-day refund.
Agentic AI changed the shape of input
For forty years, computer input meant producing the artifact yourself. The code, the email, the document: your fingers typed every character of the finished thing.
Agents inverted that. Claude Code writes the code. Codex runs the migration. Cursor edits across twelve files. What you produce now is not the artifact but the description of the artifact: what to build, what to change, what was wrong with the last attempt.
And descriptions have a different shape than artifacts. They are conversational rather than syntactic. They are long, because context and constraints matter. And they are constant, because every agent turn ends with the agent waiting for more words from you.
Keyboards were a fine tool for artifacts. For a workday made of descriptions, they are the bottleneck.
Why AI agents need voice input: the arithmetic
Here is the open math, no fake studies required. Most people type 40 to 60 words per minute. Conversational speech runs around 150 words per minute, which is why "speaking is roughly three times faster than typing" is the commonly cited ratio.
At artifact-typing volumes that ratio was a curiosity. At agent volumes it compounds. A serious agent user produces thousands of words of prompts, follow-ups, and corrections a day; we walk through a real tally in the 10,000 word day.
Run 5,000 description words through a 50 words-per-minute keyboard and that is 100 minutes of pure typing. Speak them and it is closer to 33. That hour-plus is not saved once; it is saved every working day, which is why a $99 one-time app (as of July 2026) is one of the easier purchases to justify in this category.
There is a quality argument too. Descriptions are speech-shaped: "make it like the settings page but with the sidebar collapsed" flows out loud and stalls at a keyboard. Agents parse conversational, unpolished phrasing perfectly well, filler words included. Raw dictation loses nothing.
You are a dispatcher, and dispatchers talk
Watch someone actually running three agents: Claude Code refactoring auth in one terminal, Codex grinding through a migration in another, Cursor restyling components in a third.
Their job is not typing. Their job is dispatch: review what came back, decide what happens next, issue the next instruction, move to the next station. Every high-throughput dispatch job in history (air traffic control, taxi dispatch, restaurant expo) converged on the same interface: the human voice. Eyes stay on the state of the world; the mouth issues instructions.
Agent work has the same structure, except we bolted it to an interface where issuing instructions means sitting down, finding the right window, and typing a paragraph. The dispatcher keeps getting dragged back into the typist's chair.
That is the case in one line: agents turned computer users into dispatchers, and dispatchers talk.
The loop that makes parallel agents practical
Push-to-talk dictation fixes the wordy part: hold a hotkey (in Infina, hold Option), speak, and your prompt lands typed. That alone triples your instruction speed, and for a single agent it is plenty.
But parallel agents fail on the seams, not the words. Between every review and the next instruction there are three touches: trigger the mic, press Enter, switch windows. Multiply by three agents and dozens of turns and your day is a string of tiny keyboard trips that keep you glued to the desk.
Infina's hands-free mode removes the seams. Double-tap Cmd to toggle it on (it ships off by default and is labeled experimental). Then the whole loop is speech:
- Say "type" plus your instruction: "type the tests pass, now update the changelog and open a PR". It gets typed into the focused terminal. No hotkey; "type" is itself the trigger.
- Say "send". Infina presses Enter and the agent gets to work.
- Say "open Cursor" (or "open Claude Code", "open Notes") and you are at the next station. Review, and loop.
It works from 2 to 3 feet away, so "review one, dictate to the next, send, switch" happens while you stand, stretch, or eat. The full playbook for juggling sessions is in voice prompting multiple agents, and the Claude-Code-specific walkthrough is hands-free Claude Code.
To be precise about the claim: hands-free voice control has existed for years (Talon, Apple's Voice Control). What no other dictation app completes is the whole prompt, send, and switch-apps loop hands-free in plain English. That loop is the product.
Built for prompts: on-device, raw, and owned
The agent use case shaped every other Infina decision.
On-device by default. Transcription runs on your Mac, on the Apple Neural Engine, so prompts about unreleased code never leave the machine and the loop works offline. Hands-free listening is on-device too; nothing is recorded or sent anywhere while it waits.
Raw by design. Agents do not need polished prose, so the base product optimizes for speed and fidelity instead of rewriting you. When you do want polish (emails, docs, more languages), the optional $10/month cloud add-on brings sharper cloud transcription and LLM-polished output, with a 7-day trial. That is how a $99 app beats the $15/month subscription tools at their own polish game: you own the app, and rent the polish only if you want it.
Owned, not rented. $99 one-time as of July 2026, every 1.x update included, no subscription, and a 7-day no-questions refund instead of a trial. Details on pricing.
Honest limits: Mac only, Apple Silicon required, English only in the base product, and hands-free is our newest, experimental surface with push-to-talk as the mature fallback. And if you send two short prompts a day, none of this arithmetic applies to you yet.
This is what the shift looks like
Every era of computing got the input it needed: the mouse for graphical interfaces, the touchscreen for phones. Agentic AI is the first era whose native input is description, and description's native medium is speech.
The people running three agents by voice today are not a niche; they are early. We lay out the longer arc in the voice-native era. The short version: once your job becomes telling computers what you want, the interface question answers itself.
FAQ
Why do AI agents need voice input instead of typing? Because agent work is describing work, and descriptions are long, conversational, and constant. Speech runs around 150 words per minute versus a typical 40 to 60 typed, the commonly cited three-to-one gap, so voice input turns the biggest time cost of agent work into the smallest.
Do AI agents understand rough, unpunctuated dictation? Yes. Claude Code, Codex, and Cursor parse conversational phrasing, filler words included, so raw dictation loses nothing against typed prompts. That is why Infina's base product optimizes for raw speed rather than prose polish.
How does Infina's hands-free loop actually work? Double-tap Cmd to toggle hands-free mode on (it is off by default). Then say "type" plus your instruction to have it typed, "send" to press Enter, and "open" plus an app name, like "open Claude Code", to switch sessions. It works from 2 to 3 feet away.
Can I run multiple agents this way? Yes, that is the point. Review one terminal, dictate the next instruction, say "send", say "open" plus the next session's app, and repeat, keeping two or three agents busy without touching the keyboard.
Is Infina listening to everything while hands-free is on? No. Hands-free listening runs on-device inside the app, and nothing is recorded or sent anywhere while it waits. Transcription is on-device by default too, and privacy mode is on by default, so no audio or transcripts are stored.
What does Infina cost? $99 one-time as of July 2026, every 1.x update included, with a 7-day no-questions-asked money-back guarantee. No subscription for the core app; the optional cloud add-on is $10/month with its own 7-day trial.
The bottom line
Agents did not just speed up work; they changed what your hands are for. The artifact is the agent's job now. Yours is the description, the decision, the dispatch.
Typing descriptions at 50 words per minute is running a dispatcher's job through a typist's tool. Speaking them, and sending and switching by voice too, is the version of this job that actually scales past one agent.
Infina runs that loop on a Mac for $99 once, on-device by default, risk-free for 7 days. The agents are already parallel. Your input might as well be.