Josh Johnson

Notes from JJ

Week 20October 26, 2026 · 5 min read

Do Stuff While You’re Doing Something Else

ChatGPT's computer use earns its desk, Terra joins the rotation, and your assistant gets a job.

Cream-paper editorial illustration of a hand reviewing a written plan while a gold cursor operates calendar and email windows

I started at MathCo today. New people, new systems, plenty to learn—and yes, I'm still building at home. Most of that time is going into OpenTutor. With me no longer there every day, the kids being able to navigate their digital work independently matters even more. More on the project another time. Today, the surprise was how useful my existing AI setup became before I even knew where everything was.


First Day. Extra Hands.

A real desk with a laptop, monitor, and several connected devices

Muse still dominates my everyday “go do this” AI use. But today's standout was ChatGPT's computer use: an agent operating the screens and applications in front of me.

Recently, that has mostly meant testing apps or showing the kids how to make eBay listings. Today it meant onboarding. I had two tracks: things that needed me—training, calls, actual conversations—and things that needed to get done by the end of the day. Paperwork, documentation, simple emails, getting access to the right application groups.

At points I had three or four agents working on separate tasks while I stayed with the human part of day one. Outlook was the pleasant surprise. Working in an already-authenticated session made the scheduling and administrative steps feel much quicker than I remembered. No credentials pasted into a prompt. That session still carries access, though; it's not a permission-free shortcut. Use employer-approved tools and keep sensitive work inside approved boundaries.

What impressed me wasn't a clever answer. It was how little friction there was between asking for a straightforward thing and seeing it happen in the application.

I didn't need an agent to attend my first day. I needed help with the things competing for it.


Terra Has a Very Practical Audition.

OpenAI Developers' official GPT-6.1 Terra model-page artwork

Today's standout was ChatGPT's computer use. In my experience, they're still the industry leader at it. The funny part: I did today's work with Sol, before realizing GPT-6.1 Terra had arrived. Terra's audition is next; Sol was driving today.

Still, the timing is excellent. OpenAI positions Terra for visual reasoning and precise computer use: understanding what's on a screen and taking the next useful action. That's exactly the kind of work I suddenly have more of.

The listed Standard API rates start at $1 per million input tokens and $5 per million output tokens—half Sol's corresponding rates. That's token pricing, not a flat price for finishing your inbox; context length, tools, and service settings can change the bill. It is inexpensive enough to make a small, real-world trial interesting.

My next test is less “answer this impossible question” and more “finish this ordinary workflow correctly.” Browser tasks, then some Blender work through MCP and on-screen controls. Those are different ways of driving software; I want to see where each helps.

My scorecard: Did it finish? Was the saved result right? How long did it take, what did it cost, and how often did I have to rescue it? I don't have Terra results to report yet. But if you've been waiting for a reason to try computer use on something boring and useful, this looks like a good audition—start with a low-risk task you can check.


Give Your Assistant a Job.

An office professional wearing a headset and managing a notebook and clipboard at her desk

Remember Janice? She knew the calendar, handled a little of the financials, fixed the printer, and somehow made the holiday party happen. You didn't explain her entire job every morning. She had context, recurring responsibilities, and a sense of when to come get you.

That's the leap I want more people to make with agents. Once your assistant reliably carries out a repeatable responsibility, you've given it a job. Start by saying: “Do this daily, using this background, and deliver this outcome.” An email. A text. A draft post. Name the result, not just the activity.

One of mine checks Chicago–New York flights departing two to four weeks out and emails me when a fare drops below $200. It has a route, a window, a threshold, and a reason to interrupt me. Booking could be the next step through a supported Hermes or Muse setup, with securely stored access and explicit rules for the purchase—not just a password and good intentions.

But a good assistant doesn't just stay busy. Work follows a plan. AI is good at making plans and completing small tasks; I get better results when those happen in separate runs. If the same prompt creates the plan and immediately acts on it, I haven't had a chance to decide whether the plan was any good.

Before MathCo, I split my job search into sourcing roles and working the application queue. One run found the opportunities; another worked from the reviewed list. Grants, scholarships, even finding agencies for an actor or comedian can follow that same rhythm. Ask for the separation explicitly:

Set up two weekday runs in America/Chicago. At 8 a.m., find opportunities matching my criteria and save a shortlist for review. At 10 a.m., prepare applications only for approved items, using my supplied background. Flag missing answers and email me the results. Don't submit without my approval. Share a status list so nothing gets done twice.

Separate runs or separate agents—the handoff matters more than the headcount. Save the schedules, keep the runner available, and test that execution waits for reviewed work, not just the clock.

I'm not campaigning to get rid of Janice. Even Janice deserves a Janice. Give her repetitive work somewhere to go, so she can spend more time on the things that actually need her.


The Takeaway

  • Protect the part that needs you. Give the administrative queue some help.
  • Test Terra on a real task. Check the result, not just the speed of the clicks.
  • Separate the runs. Source and plan first; execute the reviewed queue next. Say when to run and where the result should arrive.

— Josh

Quick Links