Stop Typing to Your Agent

Think about what happens when you use a typical AI coding agent. You type a detailed prompt. Wait for it to work. Read the output. Type corrections. Wait again. Your hands are on the keyboard the entire time, dedicated to managing the agent.

That defeats the purpose. The agent is supposed to give you time back, not consume it in a different way.

Voice Changes the Dynamic

When you can speak to your agent, your hands are free. You can be reviewing a design in Figma while telling the agent to fix a build error. You can be eating lunch while directing a refactor. You can be on a walk while the agent handles your email backlog.

The interaction model shifts from "sitting at your desk managing the agent" to "living your life while the agent handles tasks in the background." That's a fundamentally different value proposition.

Why It's Hard to Get Right

Voice-first interaction needs three things to work well: reliable speech-to-text, good intent parsing from natural speech, and a way to handle ambiguity without stopping everything to ask clarifying questions.

Natural speech is messy. People say "um," change direction mid-sentence, and use vague references. A voice-first agent needs to handle "fix that thing from earlier, you know, the one that was breaking" and figure out what "that thing" refers to from context.

Local speech-to-text models running on Apple Silicon are now fast enough to make this practical. You don't need to send audio to a cloud API, which solves both the latency and privacy concerns.

The Right Default

Text input should still exist for precision work - complex code snippets, exact file paths, specific configuration values. But the default interaction mode should be voice. Speak your intent, let the agent execute, check the results when you're ready.

The agents that win the daily-use battle will be the ones you talk to, not the ones you type to.

Fazm is an open source macOS AI agent. Open source on GitHub.