The First Agent: Claude Code
The whole world is seeking a paradigm.
What even ARE agents, in the context of artificial intelligence?
Everyone in the business is thinking about this at a bare minimum, and often we’re fitfully poking at our keyboards, trying to do some things. Read: something.
Anthropic clearly has the bit in their teeth in the same way GNU did with Unix way back when. Here are a couple of Anthropic guys talking stuff that now seems obvious in retrospect.
The first general purpose agent is Claude Code. Skills are the building blocks of further agents. This is the way forward …
Attention Conservation Notice:
Old man takes a break from yelling at cloud to reframe AI driven changes as an echo of the workstation wars of the late 20th century.
Skills Not Agents:
This is worth watching.
Experience:
Your First Claude Skill, on the 9th, contained my admission that six weeks had elapsed between Nate B. Jones coverage on Skills and my first getting serious with them. I’ve been running them in Claude Desktop and using it for health diagnostic stuff, since that’s where its most likely to pick up patterns of use and continuous-learning a new skill for me. I’ve been reading Nate’s ten meta-skills and could just start whipping up my own, but I really want to see the system learn my ways.
I used Antigravity for ten days, with its terrible slop sentence completion misfeature, then Google took care of that. My experience with Gemini3 is … what the Anthropic guys are doing intentionally I was doing accidentally, and a bit angrily. First, I had to switch from SQLite3 to Postgres, to stop all the tableshitting. Postgres permissions forced the model to ASK me for changes. Then I began seeing repeating patterns where it would spew out Python code. I began keeping some of those scripts, and creating prohibitions against it doing certain things, instead requiring that it ask me.
What they are doing with Claude Code happened for me in Antigravity. My Claude Code use case is … my frustration with model behavior combined with the tableshitting had led me to put tight constraints around what I had Claude Code do. Since I’ve got a computer science background, I can hold an overall solution structure in my head, and I was having the model tend to gritty details, but limited to a single area. I was also working on Parabeagle, which is a fork of a well done piece of software, so there were some implicit constraints already there.
Observations:
Some unordered thoughts from a guy who dodged programming on punch cards by just one semester …
The Claude Code /agents construct, with its own context window, contains echos of the subroutine. The prior context is “pushed” onto the “stack”, the “subroutine” runs, returning a result. This was Computer Science 111 in the 1980s.
We’ve spent a lot of time trying to improve models (CPUs), trying to expand their context (on die cache), and otherwise tweaking this general purpose execution engine in ways that seem very familiar to me, given that my first home computer had a 998 kilohertz Motorola 6502 in it.
The parallel agent coordination is the multiple cores/multiple threads paradigm we all know today, but that CPU feature only dates back to 2005.
Prior to CPUs offering paged memory management, the Intel 386 and Motorola 68030, any multiprocessing had to be cooperative, or non-preemptive. This is not to say that preemptive kernels require an MMU, the two just arrived on the desktop together. Unless your computer use began in the 1980s, you’ve never experienced this world, it ended when Windows NT and OSX arrived.
Skills are function definitions - the model’s context only holds the name and the parameters required, it gets invoked as needed.
MCP servers are …. system libraries … and remote procedure calls … and packages that implement complex data structures so that we don’t hand code things. There are a lot of familiar feels with this border between the squishy AI and imperative programming methods.
So those of us who have been around long enough have already seen this progression play out in a similar fashion.
Conclusion:
I came upon MCP servers six months after they were released. I noticed Skills immediately, but did not get to them for six weeks. If the trend continues I will catch the next fundamental advance before it is even a week old. And it’s already clear this will be a data structures thing, an improvement to how memory outside the model’s context is managed. I’m actually doing a thing involving Elasticsearch in this area, but it’s just starting to move …
What I said in Surfing The Singularity, about feeling overwhelmed … in watching that video, I could not have articulated things so neatly, but I am already paddling alongside those guys.
I’m still just as likely to drown in a double hold down shortly after New Years, but at least I’m gonna die after a short period of standing.
Coda:
Did you get your Free Perplexity? Things here are going to bifurcate in 2026. Coding in general, and agents in particular, are going to need the paid version of Claude. If you have been following this Substack for the sake of conflict tradecraft, you REALLY want that free year of Perplexity.
Perplexity’s “search first” mode, its inclusion of citations, this is the way for those doing qualitative research. They have said, but not yet delivered, that they are going to offer a way to use MCP servers from their web interface. That means that a lot of the complaints one has about ChatGPT will vanish beneath a wave of well done research that you yourself can verify by reading the citations.
There will be a lot of content on here in 2026 about Perplexity and you’ll be paying $20/month to follow along if you don’t get signed up before year end.