In 1995, my idea of "automation" was writing the right autoexec.bat so my 386SX would load the right drivers, free up enough memory, and let me boot into Wolfenstein 3D. I didn't know I was learning to orchestrate computers. I thought I was trying to play games.
Thirty years of scripting around judgement
The real trick back then was getting the serial cable config right for that first version of Warcraft. Serial-link multiplayer, two machines, two friends, one afternoon. What really blew my mind was Norton Commander over a serial cable: you could copy files from one computer to another across the wire. An early FTP, basically, years before I knew what FTP was.
Then Linux arrived, and the orchestration instinct found a real outlet. I picked up Perl, then bash, and started with the obvious things: backup scripts, update routines, the chores nobody wanted to do on a Friday. Then monitoring. Then self-healing routines that caught the same failure at 3am for the third time and just fixed it. One of the jobs that shaped me most was building an in-house deployment system for a 20-person dev team. This was on SVN, possibly CVS, years before git existed. Branch merging, dev and live pushes. A proper git workflow before "git workflow" was a phrase.
The wall that wouldn't move
Every year the scripts got smarter. But they always hit the same wall: the moment a task needed judgement, a human had to step in.
You could automate the backup, but not the decision about what to back up when disk was tight. You could automate the deploy, but not the "should we roll back?" call. Invoice processing? You could script parts of it. Not the bit where you read an email, decide if it's actually an invoice, figure out which supplier, and match it against the right account.
The wall is gone
The first time I used Claude Code, the coding agent that runs on your machine, I knew we'd crossed a line. Not prompting a chatbot. Programming an agent. Giving it tools, a memory, and the ability to decide what to do next.
I now run AI agents that do the jobs the old scripts couldn't. Vigil, my CI/CD agent, checks every repository we depend on for updates, pulls them into our codebase, resolves merge conflicts on the fly, and has a pull request ready on my desk by morning. I review, I merge, I move on. On the finance side, another agent reads my inbox, picks out the invoices, uploads them to our accounting software, and reconciles them against the billing system's records via API. That was one of the biggest energy drains in my week for years. Now it happens while I sleep.
Tools are the job now
These aren't science experiments. They're running every day. And here's what I've figured out, and it's why this feels like a career-culmination moment: building tools for AI agents is the same discipline as building tools for system automation. Same instincts, same patience for the edge cases that always eat your weekend. The agents are good, very good in fact, but they get dramatically better when you give them the right tools and the right environment to work in. That's an engineering job, not a prompt-writing job.
The Model Context Protocol (MCP) is where most of my energy goes these days. It's the clean way to build tools that agents can discover, understand, and use. Intelligent tools are where the productivity step-change lives: tools that do several related actions in one call, with good inputs and clean outputs. A dumb tool that just wraps one API endpoint is a mediocre gift to an agent. A tool that knows how to check, act, verify, and report back in one shot is a force multiplier.
Same craft, new substrate
My career was systems engineering, networking, and coding, stacked on top of each other. I spent decades learning how to make machines do more work so humans could focus on what actually needs humans. Turns out those same skills are exactly what's needed to make AI agents effective. The substrate changed — the craft didn't.
What a time to be doing this
Thirty years of scripting around the judgement wall, and the wall finally moved. The agents I run today are doing the work the 3am self-healing routines of 2005 could never reach. Same instincts, new substrate. I wish I could go back and tell my 386SX-era self what this was actually leading to.