One of many things that made Claude Code such a breakout success was Boris's spirit of embracing not what is, but rather what will be.

When he first launched it as an internal tool, it didn't work very well. I think the story goes that only 1 other person reacted to his Slack message. People were confused by the decision to use a CLI as the interface, and the models just weren't quite there yet.

But then the models got there. And in less than a year, it went from a nice coding tool to something that Wall Street believed would transform all lanes of knowledge work.

I remember when I started using Claude Code. I even surprised myself when I abandoned the IDE completely within a few days. And now like many I've been more productive as a builder than I ever have in my life, yet I haven't looked at a single line of code in months.

So what does the next paradigm shift look like?

What I'd bet on

A few things that I would bet on becoming true:

• Agents will represent the majority of internet activity -> More of a matter of "when" than "if" at this point. Lots of 2nd order effects here that I'm still digesting
• Agents will be first class citizens -> Instead of assigning a Linear ticket to an engineer who uses Claude Code, we will assign tickets to agents directly
• Agents will be fully backward compatible with current systems -> And it starts with actually good computer use

Despite things moving very quickly, Boris highlights 2 things in this interview that feel like Anthropic's version of Tesla's master plan:

1. Progressing from AI coding -> tool use -> computer use
2. Expanding from automating engineering to adjacent functions

Cognition announced Devin, the first AI software engineer, in March 2024. They were a bit early to the punch, but when they were perfectly positioned once the models caught up. And putting this all together, I just can't help but draw the trendlines and conclude that there will eventually be an AI CTO.

An AI that can replace me.

The trigger

One day I started to notice that most of the "agent orchestration" work I was doing was quite reptitive. Copying context from one place, and pasting it into another.

And then a fairly innocent question made something click for me:

So how hard is all of this now that AI can write most of the code?

It was time to find out.

An AI CTO

My thinking was to start with a co-CTO. Someone I can jam on product with at the beginning of the week, create a bunch of tickets, and just crank through them. In this barbell model, I focus on the 2 ends of the process for which there really is no "right" answer.

An if this goes well, to eventually try and create a true CTO that can fully drive the product roadmap.

What I have right now

Right now my co-CTO lives on my Mac mini and his name is Woz.

Woz can not only be assigned to Linear tickets and execute on them end-to-end, but he can also lead Linear projects. In the example below, Woz closed 100% of the tickets we needed to build usage & cost dashboards.

All I did was create an initial project with 1 spike ticket. I braindumped some considerations and assigned to Woz. Woz came back with a v1 design doc and we iterated on it a bit. V3 was solid so we created a bunch of tickets out of that.

From there I assigned Woz the first ticket, and the whole thing just works! When I approve a PR and we merge it, Woz automatically picks up the tickets that were blocked by that one.

And Woz manages all the admin work around being a CTO for me. Creating Linear tickets, closing out ones we no longer need, etc. It's great.

Woz is also all over Github.

Right now our default code review workflow waits for both Claude and Codex to finish a 1st pass review because they are good at flagging different things. Once both reviews are in, Woz will triage and decide what to commit in this PR and what to file as quick follows.

Woz isn't meant to be a black box. Woz is meant to behave like a teammate, so all the typical developer best practices still apply.

Actually in this example Woz and I had a little bit of a back-and-forth on the PR.

I had some UX feedback that required us to take a step back. Thankfully Woz is very patient and took the feedback well.

Working with Woz is pretty awesome and there's still a ton of long hanging fruit to make this better. I want Woz on Slack. I want Woz to be able to test PRs and provide evidence. Cursor recently launched a feature that does this. I'm frankly more surprised that not more people have it.

Woz has already proven to be super useful and I'm having the rest of our team adopt it as well.

I mean, how can you say no to this guy?

Follow along here for fun updates: https://x.com/wozbotai