One Year Working With AI Agents — What Actually Changed

A year ago, my Monday mornings started with an empty editor and a TODO list of features to write line by line.

Today, my Monday mornings start with a context doc — a structured brief explaining what I need built, why, and what constraints matter.

I thought AI would help me write code faster.

What actually changed was… everything else.

This isn’t an “AI replaced developers” story. It’s a story about how the role itself quietly shifted.

⌨️ I Communicate More Than Ever — Just Not Always to Humans

The funny thing is, I type more than ever — but I write less code.

Most of my keyboard time now goes into giving context, explaining intent, reviewing outputs, and refining instructions. A lot of that communication is with agents.

I find myself explaining assumptions, defining roles, describing edge cases, and iterating ideas out loud. In some ways, AI tools became a thinking partner. And that changed how I structure conversations even with people — I now give more context upfront because I’ve trained myself to do it with agents.

The faster I can articulate context, the better the results I get.

💻 Less Raw Coding, More Judgment

I don’t feel like I “code less.”

I feel like I decide more.

My daily work shifted toward reviewing generated solutions, evaluating trade-offs, spotting risks, and shaping architecture decisions.

The hard part is no longer writing lines of code. The hard part is asking: Is this the right approach?

Here’s what people miss — these models aren’t autonomous. They hallucinate confidently. They’ll generate clean-looking code that quietly breaks in edge cases you never mentioned. That’s exactly why judgment isn’t some nice-to-have soft skill. It’s the entire job now. If you blindly trust the output, you’re not moving faster — you’re shipping bugs faster.

Review isn’t optional. It’s the one step you cannot skip. Every line an agent writes must be treated like code from a junior developer who’s confident but unproven — you read it, you question it, you test it. The moment you stop reviewing is the moment AI stops being a tool and starts being a liability.

🎭 Roles Matter More Than Prompts

One big lesson: coding models perform dramatically better when you give them a role.

Instead of “Write this feature,” I now say things like:

• “Act as a security reviewer.”
• “Act as a performance engineer.”
• “Act as a skeptical architect.”

Then I run the same output through multiple perspectives. It’s less about prompting tricks — and more about structured evaluation.

🔍 Exploration Got Faster — But Decisions Became Heavier

Before, I would explore ideas slowly because implementation cost was high.

Now I can test multiple architectures, different API designs, and alternate workflows in minutes. But that speed didn’t just make me faster — it changed what productivity means.

It’s not “10x coding speed.” The real gain is decision velocity — the speed at which you can move from “I have an idea” to “I know whether this idea is worth pursuing.” I explore more, but I commit more carefully. That’s the metric that actually changed for me this year.

If you’re using coding models and feel like your workflow is changing — you’re not imagining it. You’re not coding less. You’re evolving.

And honestly? I’ve never felt more productive.

A year ago, I measured productivity by lines of code. Today, I measure it by clarity of judgment.

Originally published on LinkedIn.