Essay ·
Catching the End of the Trail
From copy-paste ChatGPT to agents, after thirty years of enterprise Oracle: what transferred, what went to landfill, and how newcomers learn good from bad.
The question I keep circling
Someone starting AI-assisted coding today has no idea how lucky and how stuffed they are, in roughly equal measure. I wonder how they learn. Not which tool to use, that changes monthly. How they learn to tell good from bad when the thing producing the code never says "are you sure you understand this?"
There is more noise in this field than in any I have worked in, and I spent thirty years in enterprise Oracle. The signal is boring. The noise has a YouTube channel.
This is the journey as I saw it from the back of the pack, catching the end of a trail that was already fading.
Copy, paste, explain it all again
My first version of this was copy and paste from ChatGPT. Ask for a function, copy it out, paste it in, fix what it got wrong, go back and ask for the next bit. Context ran out. Session restarted. You explained the whole system again from scratch, every morning, uphill, in the snow.
It was a terrible workflow and an excellent school. I was the integration layer. Every line passed through my hands on the way in, so I had to know where it went and why. When the session died I had to re-explain the system from memory, which meant I had to have the system in memory.
Nobody chose that as a teaching method. It worked anyway. The re-explaining was where the understanding lived.
What transferred, and what went to landfill
Here is the uncomfortable bit for those of us who were in early. Most of the "background of the past two years" is already landfill. Prompt incantations. Which model was best in March. The trick for getting a chatbot to stop apologising. None of it transferred.
What transferred was the thing I brought from the previous thirty years. Knowing what a system is supposed to look like. Smelling a wrong answer before running it. Having been paged at 2am by my own cleverness.
The two-year veterans are not ahead because they know AI. They are ahead because they already knew software and used AI to stop typing.
The newcomer's problem
Someone arriving now is trying to learn two crafts at once, and the tool is specifically designed to hide the first one. Code appears. It runs. The feedback loop that used to teach you, compiler errors, the hour spent staring at a null pointer, has been smoothed away. You can go a long way without forming a mental model of what you built, and nothing in the workflow tells you that is happening.
The agent is a very competent colleague who never asks whether you understood.
The noise is structural rather than accidental. Signal in this field is boring: read the diff, run it, understand the thing you shipped, do not deploy what you cannot explain. That does not sell courses. "Build a SaaS in 20 minutes with no code" does. So the loudest channels select for exactly the wrong lesson, and a newcomer cannot tell slop from substance because telling them apart needs the judgement they have not built yet. Chicken, egg, YouTube.
How do they actually learn, then? Same way we did, with the pain deferred. Something breaks in production. The agent cannot fix it because the person cannot describe it. That is day one of their real education. The ones who survive it go back and learn to read what they generated. The ones who do not blame the model and move to the next tool.
The experienced developer's steeper curve
This is the one that surprised me. I think experienced developers starting today have a harder climb than they would have had two years ago.
Two years ago the on-ramp was gentle. AI was autocomplete with ambitions. You kept your editor, your workflow, your habit of reading every line, and the tool slotted in beside you. You could adopt it without changing who you were.
Now the thing that works is agentic. You describe, it builds, you review. That does not slot in beside a workflow, it replaces one. Someone arriving today swallows 2024, 2025 and 2026 in one gulp, and the first lesson is "stop doing the thing you are best at". Not technically steeper. Emotionally steeper, which is worse.
They also arrive with a legacy combo, two legacies stacked. The codebase: written for humans, tribal knowledge in people's heads, no tests, no docs an agent can orient from, a folder structure that made sense in 2017. Agents are brilliant on a repo that explains itself and miserable on one that does not. Then the habits: control, line-by-line reading, typing as thinking. The agent produces faster than an expert can read with full attention, and the expert's instinct is to slow the agent down rather than change how they review.
A newcomer with a greenfield repo and no habits sidesteps both. That is the one genuine advantage they have, and it is not nothing.
But the window for coding from scratch did not close. It is wider than it has ever been. What closed is the window for easing in. The experienced developer's real handicap is that the part of the job they identified with, producing code, became the cheap part, and the part they often did reluctantly, specifying, reviewing, deciding, became the whole job. The ones who accept that trade are more dangerous than any newcomer, because they can tell when the output is wrong. The ones who do not sit with the legacy combo and call it a window closing.
Which clock you read
I thought I was early. Depends which clock.
For AI coding in general, two years ago was not early. Copilot had been in production since 2021, ChatGPT since late 2022, and by the end of 2024 Cursor was already the editor everyone argued about. People had been pasting stack traces into chatbots for two full years. Mid-period, not frontier.
For agentic coding, the thing that actually changed how I work, two years ago was prehistory. Claude Code did not exist until early 2025. "Let it edit the files" was a novelty people tried and mostly turned off. The default was still: you drive, the model suggests.
So I was early to the second thing and roughly on time for the first. That is the honest version of most early-adopter claims in this field. The ground keeps moving, so everyone was early to something and late to something else at the same moment. The people who sound earliest are usually the ones who switched clocks most recently.
A word on Copilot
I was never going to use Copilot. It's Microsoft. Thirty years of Oracle will do that to a person, and I accept this is a religion rather than an argument.
But there is a product-shape point underneath the prejudice. Copilot sits inside your typing and finishes your sentences. For someone who thinks by typing, that is a colleague leaning over and pressing keys. Chat, even clunky chat, kept the division clean: I decide, it drafts, I integrate. That is the same division I ended up with in the agentic version, just with the drafting bit grown up. I went from the ugliest version of the right model to the good version of the right model and skipped the polished version of the wrong one.
And Copilot did work, on the balance sheet. Microsoft built the thing that fits inside the existing workflow because that is what enterprises buy. Nobody gets fired for autocomplete. The company with the most to lose from developers changing how they work built the tool least likely to change how they work. Not a conspiracy. Gravity.
For a solo shop, the only balance sheet that counts is your own.
It is a fun journey, though
Thirty years of waiting for the tools to catch up with the ideas, and then they did, all at once, in the last stretch of a career. Most people in my position got the opposite timing: the good tools arrived after they had stopped building. I got them while I still had the itch and the scars to use them properly.
So if you are coming in now, with the noise at full volume and no copy-paste era to teach you the hard way: read what the machine wrote. Run it. Understand the thing you shipped. It is boring advice. It is also the only advice that will still be true in 2036, when the copy-paste era is a story nobody believes and we are all complaining about something else entirely.