Benjamin Lannon

Select a theme. Click on the overlay or the button again to exit

How I use AI (Fall 2026)

Posted on:

A few months have passed since my last "How I use AI" post anad a good chunk has changed in said months. Lucky for me, still only $20 a month is going into a codex sub, but how I interact with AI and also where I do that work has changed.

The Where

My main interface of choice as of a few months back has become T3 Code where it is not a harness like Claude Code or Codex, but rather a frontend you run in front of those harnesses. So I just have to login into the Codex CLI on my machine and then open T3 Code and I can use Codex across my apps with such. And say 3 months down the road if I switch subs, I don't have to change the interface I use.

Of note, the remote capability I feel is something that is starting to open some doorways into a new style of development where the instance I am interacting with may not be on the same machine as the one I have T3 Code open on. So I could open T3 Code on my MacBook Air and create a new thread on a project on my Framework Desktop and the agent itself would be working on the Framework while I orchestrate via my mac. Alongside this, this also can be done similarly from my iPhone via the T3 Code mobile app. I feel for the longest time doing dev work from a phone or tablet was something that was interesting but never really grabbed me. That said, T3 Code is just a chat thread and it works suprisingly well. On top of such, since my phone is on my tailscale network, I can leave the house and continue on, and particularly other apps that crumble in spotty connections, T3 Code still stays rather responsive in those situations.

The How

Whenever I am starting a new idea and don't have ane existing project or just want to make a one-off thread, I have a folder ony my computer called Playground where I just spin up threads if like if I am in ChatGPT. That said, I feel the codex harness is a lot better than the standard web app for ChatGPT as if I want to add some skill or export an artifact, the folder is there for those inputs / outputs.

For scripts I want to run on a schedule, I made node-schedule with GPT 5.6 Sol which will deploy a nodejs script to my local Kubernetes Cluster without having to build a docker container for every script. Similarly, T3 Code as of this week now has a scheduled task tool as well so both avenues offer a similar functionality and I just have to decide whether I want a prompt in a coding agent to trigger it or just have it trigger some javascript directly.

For one thing I am doing with this scheduled task concept, I wrote previously about discord webhooks and the following is what I'm using it for so far (they all aren't fully standardized on one tool or another so keep that in mind):

As well, an agent skill I've been going to for exploring a new idea is Matt Pockock's "Grill Me" skill, particularly grilling. Primarially use it for dev work but have been thinking of using it for non-dev work into the future. Otherwise I tend to stay fairly barebones with my agent environment otherwise.

Finally, in terms of restrictions on the harness, Codex's "Auto Mode" which spins up a subagent when doing tasks to do a check if the command it will run is safe and it has worked reasonably. Much better than one of the alternatives having to approve every read or bash command.

Remaining Paper Cuts

The tooling is continuing to get better and better, but it is still not perfect. Some of these paper cuts are about AI agents in general rather than just coding agents. Of note: