The voice-first glossary.
The vocabulary of voice-first computing, defined by the team building it. Each term links to a deeper piece on how it actually works.
Voice-first
Voice-first means voice is the primary way you interact with software, not a feature added to a screen-first product. A voice-first product organizes itself around speaking and listening: deliberate capture, spoken intent, and output that goes where the work is. On a work computer that takes the form of a voice agent that acts across your apps, not just a microphone inside one app.
Read more →AI voice agent
An AI voice agent is software you direct with your voice to complete real tasks in real systems. It listens and talks, resolves what you mean in the context of your actual apps and files, and executes the steps through your apps' own APIs, with a person confirming the consequential ones. The result is a changed system, not text on a screen. It differs from a call-center voicebot (which serves a business's customers), from dictation (which types), and from a single-shot voice assistant.
Read more →Voice-to-action
Voice-to-action, also called speech-to-action, is the design where spoken language produces executed actions rather than text: the agent resolves your words to real entities in your workspace, chains the steps across apps through their APIs, and confirms the consequential writes with you. Dictation converts speech into text and stops at the cursor; voice-to-action starts where dictation stops.
Read more →Background agent
A background agent is an AI agent running a delegated, self-contained task on its own multi-step loop, separately from the conversation that spawned it. Inside a voice product the delegated instance is called a sub-agent: you hand it a bounded task by voice, keep talking, and check in later by describing the task rather than tracking an id. In mrmr, reads run while it works and every write pauses the run for your approval; stopping discards any pending write.
Read more →Human-in-the-loop
Human-in-the-loop means a person stays part of the decision cycle: the agent proposes, and a human approves, edits, or rejects before the action takes effect. For agents that can act on your behalf it means nothing consequential happens without a person authorizing it. Good human-in-the-loop design asks about the right things (consequential writes), not the most things, and enforces the asking somewhere the model cannot overrule.
Read more →Approval fatigue
Approval fatigue is what happens when an AI agent asks for confirmation on too many actions, including trivial ones. The requests blur together, you stop reading them, and you start approving reflexively, which turns the safeguard into theater. The fix is calibration: reads and genuinely low-risk reversible actions skip the gate so the confirmations that remain are worth reading.
Read more →Write gate
A write gate is an enforcement layer, below the model, that refuses an agent's write action until the user has approved a confirmation card matching that exact action and its exact arguments. Reads run freely, consequential writes confirm, and unrecognized actions default to asking. Because the gate lives in the application rather than in the model's instructions, injected content or a misbehaving model can propose a different action, but cannot execute anything you did not approve.
Read more →Prompt injection
Prompt injection is an attack where instructions hidden in the content an AI system reads change what it does. Direct injection is typed at the model by whoever is talking to it; indirect injection arrives inside content the agent processes on your behalf, such as a web page it fetches or a document it summarizes. It cannot be fully prevented today, which is why the defense is layered: least-privilege tools, gates enforced below the model, and fail-closed defaults.
Read more →Voice dictation
Voice dictation converts speech into text at your cursor. It is a genuinely useful tool for turning speech into writing, and its boundary is the point: it cannot send the message, file the ticket, or move the meeting. Everything after the words land is still manual. A voice agent is the layer that acts after dictation stops.
Read more →Wake word
A wake word keeps the microphone always listening for a trigger phrase ("Hey Siri"). Push-to-talk opens the microphone only while you deliberately hold a key or button, so nothing is captured between commands. Push-to-talk is the flow- and privacy-first choice for work: the microphone exists exactly as long as your intention does, and a hands-free mode can be latched on demand without ever removing the off state.
Read more →On-device vs cloud
On-device processing keeps your audio and data on your machine; cloud processing sends them to third-party AI providers in real time. Local-first wins on data never leaving the device; cloud wins on capability: accuracy across languages, and the reasoning quality an agent needs to act across your work apps. The honest answer is per workload: local for content that must never leave, a well-architected cloud tool for everyday work content.
Read more →Screen context
Screen context is an agent capability where, given an explicit request that refers to what is on screen, it captures an image of the front window and processes it through a vision model to answer the question ("what does this error mean?"). It should be permission-gated (an opt-in macOS Screen Recording grant), limited to the single request, and not retained after it is answered.
Read more →Per-app writing style
A per-app writing style is a setting that adjusts dictated text to the register of the app it lands in: formal in email, casual in personal chat, structured in docs. It exists because a perfectly transcribed message in the wrong tone is a correction waiting to happen, and hand-adjusting tone after every dictation is a small tax that compounds.
Read more →Switching tax
The switching tax is the attention cost of leaving the app you are working in to do a small task elsewhere and returning. Research from UC Irvine found interruptions take an average of 23 minutes and 15 seconds to fully refocus from. Knowledge workers switch apps hundreds of times a day, and voice-to-action exists to make those small cross-app tasks cost seconds instead of switches.
Read more →Get private beta access
Book a short setup call or join the invite list for Agent Mode access.
