AGI is already here (sort of)

My Non-Hype take on Claude Code (It's amazing for non-coders)

Originally published on Substack

What is Claude Code?

Claude Code is Anthropic's desktop applications (also command line tool but don't worry about that) that lets you delegate complex tasks. Unlike chat-based AI models where you're essentially having a conversation, Claude Code can actually do things, it can read and write files, execute commands, browse the web, interact with APIs, and chain together complex workflows using all of these. It can also build tools to do things that it can't currently do (it's a slow singularity of sorts).

The difference is fundamental. Chat models are like having a very smart advisor who can only talk. Claude Code is like having a junior employee who can do stuff. It uses MCPs (Model Context Protocol servers, basically plugins that let Claude connect to Gmail, Slack, databases, APIs, whatever you need and allow it to do things like send an email) and Skills (instruction files that teach it how to do specific tasks your way, and also how to use multiple tools).


How I Use It (And It's Only Been Two Weeks)

I don't primarily use it to code I use it for "business" work, and it's amazing, with a bunch of skills set up and MCPs I use it to:

CRM & Sales:

Internal Operations:

Client & Compliance:

Content & Web:

Task Automation:

Oh, and it's really good at coding too.


Why This Is a Game Changer

These all sound like small tasks individually, but collectively, it's a fucking game changer. What I used to apportion a week to now takes two days.

What makes it powerful from a business perspective isn't just all its MCPs and interconnected tooling, it's that I can sit in one interface, create 10 tasks in parallel, and it's like delegating to a team. Yes, I have to review the output, but it gets it right most of the time.

And here's what's really cool: when it gets something wrong, I tell it why and ask it to try again. When it then hopefully gets it right, I tell it to update its Skills files so it doesn't make the same mistake twice. The system learns and improves the more you use it.


What Makes It Work

Claude Code works best when two things are true:

  1. It's grounded with really good data. Incidentally this is also why our digital twins at Brox work well, garbage in, garbage out applies to everything.
  2. You invest time describing your workflows. How your files are structured, how you want things done, what your standards are. You also need a clear goal, not necessarily fully developed, but you have to know enough to know what you want.

The Implications

Are jobs fucked? Yes and no

I can't quite see it replacing whole functions yet, but I can see that one person can now do a whole lot more. We'll need fewer people for each function. The world is going to split between people who can use these tools well, probably those with the ability to think abstractly and break down complex problems, and those who can't, or who can only use them at "Level 1." (This is why I'm investing a lot of my time getting my nontechnical team using Claude Code, they are doing great)

This will cause unemployment for many, especially older workers who have a less flexible model of what cognitive work looks like, but the young and smart this should be a good thing (unless AI kills us all or replaces the economy, in which case we're all shit creek, even the billionaires in their supposed bunkers).


The Future of Software

More broadly, I think most SaaS software is fucked.

Sure, it will take a while to displace Salesforce and its ilk. But not forever. Every CRM, every workflow application, accounting software, sales software, cap table management software, eventually even Microsoft Office, it really is easier to create documents with HTML, or use Python and a CSV file instead of Excel.

Which Businesses Survive?

The companies that remain (I think) will be those with:


Final Thoughts

Claude Code isn't perfect (there are other tools as well like GPT Codex). It makes mistakes. It sometimes confidently does the wrong thing. But the feedback loop: use it, correct it, update the skills, use it again, means it gets better over time.

Lastly, I believe that the future of software will be the MCP. Every piece of software you can't build yourself will become an MCP for agents (mostly human-supervised I hope) to interact with. If you're in software and your business doesn't work as an MCP then you are double screwed.

Also Claude have released a feature called Co-work which is a hybrid of Claude Code, it's ok but it's annoying for a lot of reasons, this will probably change.