Taking coding agents for a walk

October 17, 2026 3 min read

I am a huge fan of controlling agents by voice, and talking to AI in general, as if talking to a real person. OpenAI's Advanced Voice Mode and the more recent GPT-Live were both mind-blowing the first time I used them. One thing that makes GPT-Live more special is its ability to call tools. It can search the web to look things up and check facts as you talk. It's basically amazing for all kinds of questions, brainstorming, and language practice. It works really nicely when taking a walk, for example, when you'd rather speak and listen than type on your phone.

However, I've wanted more tools available in voice mode. For example, I'd like it to look through my projects so it has more context when we're discussing a design. As I'm writing this, I'm noticing that this may have improved just recently, as voice mode seems to be able to run Python and read my emails because I have Gmail connected.

But what I recently discovered is that you can actually use Codex on your Mac by talking to your phone, through the ChatGPT Mac app. This to me feels insanely powerful, because it means you can remotely control your computer using voice alone. (I really couldn't care less about opening a terminal on my phone to do stuff.)

ChatGPT voice session on an iPhone running bash commands to check the Mac’s uptime and reporting 24 days, 21 hours, and 45 minutes. — Checking my Mac's uptime from voice mode on my phone. And yes, you literally need to mention bash, otherwise it doesn't realize it can do this. Something they could improve.

Yesterday, my Claude Code weekly usage was about to reset, and I still had some quota left. I figured I'd put it to use by planning and starting an implementation while out for a walk.

I also asked it to get the design reviewed with consult-llm, my CLI for consulting other models. It actually read the consult skill, called the CLI to have Astra and Opus review the plan, and came back with their feedback. Getting that back in a voice conversation while out walking was really cool.

Once the plan was ready, I told it to finish off by starting Claude Code in a new tmux session and using workmux's coordinator skill to implement it phase by phase, with a separate worktree for each phase. The whole thing worked near perfectly, apart from a few disconnects during the call.

This again felt like one of those mind-blown moments you occasionally get with AI, when you see it do something you hadn't quite realized was possible.