The AI Crutch: What Happened When My Agent Mode Turned Off

The Discovery of Agent Mode

I'm coming up on a year since I started my first job as a software developer.

When I started, I didn't know what "agent mode" was. Like most beginners I asked ChatGPT for syntax and pasted snippets in where I needed them. Then somewhere in the middle of those months I found agent mode, and everything sped up. I was shipping features fast and closing PRs by the hundred. The code wasn't great, looking back. But for a junior it was good enough to pass.

The Wake-Up Call

My company gives us a Cursor subscription with a fixed pool of AI credits each month, and it doesn't renew until the month is up. I burned through mine early. I'd been leaning on the agent to write and think through basically every feature.

Then one afternoon, credits gone, I had a feature due that day. The TRD was already written — with AI, naturally — so I wasn't worried. It had been a year. I'm a software engineer now. How hard could it be?

I opened VS Code, opened the codebase, and just... stopped.

Five minutes staring at the screen. I didn't know where to start. The project I'd contributed to most felt like something I'd been added to that morning. I couldn't tell you where the logic lived. I couldn't tell you where the feature was supposed to go.

That's when it landed: I'd lost the ability to think through and write code on my own.

A Shared Reality

I wasn't the only one.

Months earlier a coworker had written a post about feeling like he wasn't learning anything anymore — just shipping. I ignored it. I'd been feeling something similar, that one of my dream jobs had quietly turned into a chore, but I brushed that off too. Then a few days after my own afternoon, another colleague hit exactly the same wall.

That coworker's post, "I'm Not Sure I'm Learning Anymore", is worth reading in full. He put into words the exact unease I brushed off months before it caught up with me, and it reads very differently now than it did the first time. If any of this resonates with you, his side of it is the other half of the story.

So yes, we got lazy. But the tool was also handed to us with the pressure to deliver already turned up. Taking a day to properly understand a feature, or map the architecture, or write the core logic by hand, felt indefensible when the agent could produce something in seconds. Every incentive pointed the same direction. We followed it.

The Path to Recovery

We eventually sat down as a team and talked about it, which helped more than I expected it to. The damage is done, though, and undoing it is slow and genuinely unpleasant.

These days I go through each feature deliberately. I write my TRDs with real input from me instead of prompting my way to one. I've had to relearn our codebase more or less from scratch, which is a humbling thing to admit about code that has my name all over it.

I haven't dropped the agent. The industry is going this direction and refusing to use it would just put me behind. But the workflow is different now: I read every line it produces, and anything I don't fully understand gets flagged and picked apart until I do.

It's slower. I'm also writing code again instead of just generating it, and I'd take that trade every time.