How AI in November 2026

It’s November 2026 and I’m not sure how to edit a file, but I have made enourmous progress on administrative backlog.

GrokBot is acting as my personal chief of staff, operations manager at my company, and “Mrs Hudson” has filled out and managed more school forms, bizarre ed-tech systems and familiy calendars than you can believe. It managed the hotel reservation and airport shuttle for this last minute flights (though I was unimpressed with the flight choices.)

For larger build projects it spins up a Cursor environment, which I used to build out not just a site for the new gym I’m trying to start up but also I had it put together a proposal for the planning board to get a change of use on the building.

For deeper work I’m leaning more on Amp — this plugs into my ChatGPT, Claude and XAI subscriptions and uses whatever models it feels is the right one. I have some opinions, but I lean more towards “the high thinking settings” in the beginning and then the “workhorse” settings when executing. Amp is pushing the limits of the harness more than anything right now.

Off the laptop: Cloud, Orbs, Remote Compute

Did I make a mistake with this awesome laptop that I own? Is it time for the new folding phone form factor to replace everything, with just a large keyboard and monitor at special work-pods and then we can be free of the laptop? How long have we been saying Network is the Computer, and is the time now?

Agents have morphed into useful things, something like a personal assitant in the sky. Instead of letting Openclaw run ruffshod over your harddrive — or better yet hermes agent — security and isolation have been solved by wrapping up everything into a cloud computer with gated access, where you are making connections and adding plugins rather than attaching mcp servers or installing skills. It just works, it works while dictating into the phone, it works when you get back to your big screen, it works off somewhere in the background, presumably in some rural data center.

More and more I want it off of my computer; or I want to run the whole thing myself on the computer. But for these stomping around the internet tasks, I want to dictate and walk away, check in on it, or have it check in on me.

People to follow

Three people I keep reading when I want to remember that the interesting work is not the model card.

Thorsten Ball — stretching the harness

Thorsten is pushing what an agentic coding harness can actually do. Amp’s orbs, the disposable scaffolding, the “give me irrefutable proof” loop — he’s writing the frontier practice in public while shipping the product. When I want the harness to go further than a chat pane on my machine, I look at what he’s publishing.

Lauren Tan — pstack as craft

Lauren put the wisdom of a genius developer into pstack: playbooks, principles, /poteto-mode. Not more code — less code, verified. It’s the opposite of slop throughput. If Thorsten is stretching the harness, Lauren is stretching the judgment you put inside it. (She’s also on the AIE stage talking about slop vs craft — the timing is not a coincidence.)

Matt Pocock — best practices for everyone

Matt keeps doing the unglamorous thing: communicating software best practices so everyone can use them. Skills, specs, tickets, the deletion test. He takes what the best practitioners already know and turns it into something you can install and run — see skills v1.1 and the skill hell writeup from the fair.

Domain specific Harnesses

I care less and less about the models, and more about harness that are focused upon the workflow and usecare. GrokBot is my main switching hub, and it hands a bunch of stuff of to cursor for more work. Amp is great, and that plugs into my OpenAI, Anthropic, and XAI subscriptions. I’ve been using Antigravity a bit for a specific client, and like it, and the DeepSeek Harness is amazing. I even wrote a plugin.

I really like Claude Desktop, but I don’t want to be tied to claude. (I still don’t understand how the codex app is supposed to work.). The ability to swtich between models is more important than the models themselves.

…but they need to work in the cloud, and I should be able to seamlessly interact with them over the phone and the laptop.

Personal Assistants

If we solved coding back in December, somewhere recently we’ve solved for personal assitants. This is broadly useful for everyone. I haven’t played with Muse, and am very happy with GrokBot tweaking my hotel schedule and making phone calls for me. OpenAI has Dots, which I haven’t tried, and Claude has something, these improved and elevated Openclaws that just seem to work. I expect this to be the way that most regular people end up using it, the AI for the consumer outside of the more generic “chat bots”.