We Are All Agentic Engineers Now

October 23, 2026

By Jordan Last, with AI. Written by AI from my own words in episode 1 of the Millillion podcast, using these skills.

I’m a software developer. I started my undergraduate degree in 2009, graduated in 2017 with a bachelor’s in computer science, and I’ve been developing professionally since about 2013. I love technology. I believe it’s one of the key forces that brings prosperity to the world and helps end physical human suffering. It isn’t without its weaknesses, but the benefits far outweigh the costs.

These days I don’t call myself a software engineer anymore. I call myself an agentic engineer with an emphasis in software development. And I believe that anyone in a STEM field, maybe anyone in any field, is an agentic engineer now.

From chatbots to agents

An agent is an instance of an AI that can actually go out and perform work for you, somewhat autonomously, and sometimes very, very autonomously. It’s the evolution of the chatbot.

Since around 2022, with the arrival of ChatGPT, chatbots have been everywhere. That’s a back-and-forth: you ask a question and get a response, a lot like searching the web. And for a long time those answers lied to us constantly. They would make up CLI commands and API endpoints. You’d think, “I can’t believe the exact command I need for this exact situation exists,” and then you’d try it, and it definitely did not exist.

I first tried GPT-3 in 2020, and like everyone, I thought it was incredible. I had an edtech startup at the time and wondered if it could generate practice problems and assessments. We quickly found out it just wasn’t good enough. Then ChatGPT came out, then GPT-3.5, and if you used these things at all you could feel the jump in intelligence each time.

Late 2024 is when I first experienced real AI agents. Instead of pasting a complicated Rust error into a chat window, I could tell the AI, right in my editor, to actually write the code. It kind of worked. It worked just enough that I obsessed over it. I spent a couple of weeks in early 2025 trying to force an agent to follow my plans, writing what was basically a skill before there was a skill standard: a markdown file with a process to follow, step by step. It would go off the rails every time. It just wasn’t ready.

But I could feel it coming.

The threshold

Through 2025 the agents kept improving, but my team and I were still really coding by hand, reviewing everything piece by piece. Then around October and November 2025, with models like GPT-5.1 and Opus 4.5, came the crazy jump in capability. That’s when I realized: I think these things can generate all of the code now.

So that’s what I started doing. In December 2025 I used an AI to do a very complicated tax return for the first time. I take taxes extremely seriously, so it took me 24 straight hours of building verification schemes and generating code to make sure every number was correct. I felt good enough to sign it.

Then I joined a construction startup and took on a complete AI transformation. We replaced the manufacturing software we were going to use with our own, and I vibe-coded that demo in a day.

Generating is easy. Verifying is the job.

Eventually that software had to go into production. People needed to use it, and the company needed protecting from hacks and data leaks. So I put my blockchain hat back on, where things have to be as close to perfect as possible, and dug into all that generated code. I didn’t feel comfortable with it.

I still believe that generating 100% of the code with AI is the right thing to do. The question is how you verify it.

That’s the heart of agentic engineering. The AI generates a stream of evidence about the work it’s done. Do you really want to inspect all of that work meticulously with your own eyes? Sometimes you have to. But more and more, you want to judge the evidence the AI presents to you.

For example, I have a skill for abstractness, a very important property of good software. I can tell the AI to make sure a PR or a file is sufficiently abstract, and it reasons through the code and changes it. I don’t want to review every line it touched. I want to trust that the skill is correct, and trust that if the AI tells me it ran the skill, the code is now sufficiently abstract.

That’s the goal: keep raising the level of abstraction you work at, and give the AI ever bigger, more ambitious tasks. You test and probe, you judge the evidence, and you gain confidence. In the beginning you’ll vibe things and check a lot of the work afterwards. Once you have your own pipeline that refines slop into clean code, whatever you’re applying it to, you can start to trust it.

Jagged intelligence, and why your expertise matters (for now)

These AIs are jagged intelligences. Picture a bar chart of different areas: software engineering, mathematics, accounting, art, teaching, writing. The AI isn’t equally capable at all of them, so the chart is jagged, with gaps. Humans can fill in those gaps, and that’s why we’re still economically useful for the moment.

That’s also why it’s agentic engineering with an emphasis. My emphasis is software development, and that’s what gives me the ability to judge the output and the verification evidence the AI presents. Me and the AIs together can move far faster than before.

I believe we will reach AGI, and that at some point the software engineering I do today will be completely automated: so good, so verified, formally verified even, that maybe I’ll just review specifications and behavior. I’m not necessarily scared of that. I don’t believe being economically useful is the purpose of humanity, and I think we’re headed toward outrageous amounts of abundance. But for now, our skill sets are still vital as judges.

Training the next generation

My team is about to grow from one person to two. Our new developer is fresh out of school with a computer science degree. He learned in the AI era and has barely written code by hand. What am I going to teach him? Maybe he doesn’t need my full skill set. Maybe he uses the pipeline I’ve built, and his real job is to understand the office, the factory and the construction site, hold the whole system in his head, and make sure the experience is beautiful and works well.

It’s a new world. But whatever he ends up doing, he’s going to be an agentic engineer: managing agents, writing skills, and thinking critically enough to judge.

Fully embrace AI agents, wisely and deliberately. Become an agentic engineer.

— By Jordan Last, with AI