Craig Hewitt, founder of podcast hosting platform Castos, appeared on the Zero to Agent in 30 Minutes podcast to demonstrate how he uses the Codex application’s voice mode to operate parts of his development workflow without touching the keyboard. He explained how voice mode differs from plain dictation tools and walked through how he controls his browser and other local applications with conversational instructions.
Setup and how he runs a hands-free session
-
Enable browser and computer access: In the Codex app you must enable computer use and configure browser access so the agent can interact with websites and other applications. Craig recommended requiring explicit approval before the agent accesses most apps or sites.
-
Start a voice session: Launch voice mode and give the agent instructions conversationally. Craig noted that the voice interface stays available as you switch among applications, so you don’t need to keep returning to the chat UI.
-
Assign browser tasks: He had the agent open websites, search for information, and navigate pages. He also used browser control for routine tasks like filling forms when the agent already had the required context.
-
Let the agent work across applications: Computer use extends the workflow beyond the browser. In Craig’s demo the agent opened Cursor, located a specific repository, reported on uncommitted changes, and then committed those changes after receiving permission.
-
Add specific page content to the conversation: Craig showed how selecting part of a web page and adding it to the chat gives the agent the context to act on a particular element, such as a section of an interface you want to change.
Risks and permission management
Craig emphasized keeping permissions narrow because browser and computer control introduce real risks, including unintended actions and prompt injection from web content. To manage these risks he:
- requires approval for most applications,
- grants broader access only to selected tools and local development sites,
- avoids untrusted sites,
- and uses human review for decisions.
Craig’s experience shows that voice mode can let you direct browser, application, and some coding tasks conversationally, but the demo also highlighted a delegation question: once an agent can act on your behalf, how much are you comfortable handing off? His approach is to tightly control permissions and require human consent for sensitive actions.
Next episode
In the next Zero to Agent in 30 Minutes episode, Jayeeta Putatunda, forward deployed AI engineering lead at Turing, will build an agent to help financial analysts keep up with a constant stream of new information. She will demonstrate how the agent categorizes financial news, ranks stories against analysts’ coverage profiles, and explains why each development may deserve attention.
