Two weeks ago I wrote that I could finally talk to my work agents through ChatGPT's live voice mode. I expected that to be what would finally get me working while walking: phone in my pocket, AirPods in, chatting with my agents.
That hasn't really happened. I do work from my phone a lot more now, but for a completely different reason.
Voice-only is too slow
A screen turns out to matter much more than I thought. One glance at my phone tells me which agents are done, which are still busy and where something went wrong. Having a voice read all of that out takes ages, even when it talks fast, because it can only tell me one thing at a time.
The voice agent also gets things wrong. There's a powerful model underneath, but somewhere between me talking and the coding agent receiving the task, information gets lost. I regularly got answers that were off or just irrelevant, for tasks the coding agent itself would probably have handled fine. Putting a chatty middleman in the chain didn't make anything more reliable.
What I actually want is to speak my instructions and read the results.
T3 Code
The thing that did change my workflow is T3 Code. It's an open-source interface for existing coding agent CLIs like Cursor CLI, Codex and Claude Code. It isn't a model or an agent itself (the project calls it an "agent harness control surface"). There's a desktop app, a web interface and apps for iOS and Android.
I run it on my Linux server, the same machine where my projects, VPNs, Docker and dev tools already live. I start a few tasks from my laptop, close the lid, and the agents just keep going. Later I open the same threads on my phone, check what's finished and give new instructions. Back at my laptop, I continue in the same threads. Nothing needs to be handed over.
This is what I was missing. With Cursor on my laptop, closing the laptop meant losing my grip on the session, and I couldn't pick it up on my phone. Cursor's cloud agents do run remotely, but not on my server, so they don't have my files, internal network and tools. And SSH from my laptop works fine for files and processes, but it doesn't give me one agent session I can share between devices.
T3 Code is also good at handling a lot of threads. I archive a thread when I'm done with it, snooze one so it comes back later, and threads get settled automatically once their PR is merged. I treat my thread list like an inbox and try to keep it close to empty.
Subscriptions instead of API prices
Because T3 Code drives the official CLIs, all of this runs on my Cursor Ultra and ChatGPT subscriptions (the ChatGPT one is about $100 a month). I used to do something similar with the Pi coding agent so I wouldn't be tied to one provider, but T3 Code works much better for me on both desktop and phone.
My rough guess is that my usage would cost somewhere around €4,000 to €5,000 a month at regular API prices. That's my own back-of-the-envelope number and it depends a lot on models and caching, but I'm usually nowhere near my subscription limits. Of course the providers can change those limits whenever they like.
Dictation instead of conversation
The other half of the story is VoiceInk with NVIDIA's Parakeet V3 model running locally. Text shows up about as fast as I can talk, it's very accurate in my experience, and it works without an internet connection. I now use it for almost everything I type, at my desk and on my phone.
This is the voice part that works for me. Nobody in between has to interpret what I meant; the agent just gets my words as text. Agentic coding is mostly writing instructions and reading results anyway, so a tiny keyboard is barely a problem anymore. For a lot of tasks my phone hardly feels slower than my laptop.
So, am I walking more?
A bit, but not as much as I hoped. I thought I had to get rid of the screen to work while walking. It turned out I mainly had to stop depending on my laptop. The thing that kept me at my desk is mostly gone now, so whether I actually go outside more is up to me.