I stopped writing code. Sort of.

For months, I kept hearing about agentic AI tools but didn’t actually try them. I was using ChatGPT and Claude daily for drafting docs and rubber-ducking problems, but letting an AI write code, run commands, and open PRs? That felt like a different thing entirely.

I’d spent years learning how the sausage is made. Parsers, ASTs, control flow, all the stuff that made me feel like I understood what was happening when I wrote code. The idea that someone could describe what they wanted in plain English and get working code back was hard to square with that.

Figuring out what was real

The AI discourse is noisy. CEOs had been predicting the end of software engineering for over a year, and most of those takes seemed designed to sell something. I was skeptical.

But then I started noticing people I actually trust changing their tune. Kent Beck, who created TDD and has been programming for five decades, was talking about being re-energized by AI agents. Martin Fowler was on a podcast discussing how AI changes software architecture. Gergely Orosz was dedicating entire issues to agentic coding. Addy Osmani had written a whole book about it. Simon Willison was prompting Claude from his phone to create PRs on his open-source projects. These weren’t hype merchants. They were people whose work had shaped how I think about software. That got my attention.

Actually trying it

The timing worked out well. I’d just joined Instacart, and within my first few weeks I noticed most engineers had already moved away from Cursor to Claude Code. My onboarding buddy was direct about it: ditch Cursor and go all-in. That matched what I’d been seeing externally, so I did.

Going back to the terminal felt strange at first. It reminded me of what programming must have looked like in the 90s before GUIs took over. There was something oddly satisfying about it.

What surprised me

The first real win was codebase exploration. I was dropped into a deeply complex, domain-driven Ruby codebase with dozens of active contributors and years of history. I’ve joined new codebases before, but this was different. I was able to orient myself faster than I ever had. Not just skimming files, but actually understanding how pieces fit together.

Simple Jira tickets my mentor handed me early on could be handled in a single prompt. That part was almost jarring.

What I didn’t expect was where the difficulty actually showed up. It wasn’t the code. It was making sure I understood what Claude had done so I could genuinely own the PR I was putting out. The bottleneck shifted from writing to understanding and directing.

What actually changed

I’d been equating my value with my ability to type code, remember syntax, hold a mental model of a codebase in my head. Some of that is less important now.

But the things that actually matter, understanding what to build, knowing when something is wrong, designing systems that hold up under real-world pressure, asking the right questions—none of that went away. If anything, it got more important.

Within a couple months, I went from copy-pasting between ChatGPT, Copilot, and my IDE to working more like a director than a typist. I’m not just writing code anymore, I’m directing it.

My current setup

My workflow looks nothing like what it did before. I run tmux with nested sessions: each outer pane is a project, and inside each one I have Claude Code running in one pane and everything else in another. lazygit for version control, Vim for quick edits without context-switching. I’ll sometimes go an entire day without opening my IDE. MCP connected to everything I can.

Addy Osmani wrote about this shift in Death of the IDE?, the idea that the center of developer work is moving away from the editor and toward supervising agents. That matches what I’ve been experiencing.

I’m also skeptical about whether this is the right setup long-term. I’ve read enough Cal Newport to know that humans are bad at multitasking, even when we think we’re not. There’s probably a limit to how many sessions you can actually manage in parallel, and I haven’t found it yet.

The transition wasn’t smooth. Some days I felt like I was losing something. More days I felt like I’d figured out something new. But honestly, I’m still figuring it out.