Three months ago, I stopped writing code entirely. Not because I was promoted, sidelined, or burned out—I still deliver features, solve architecture problems, and review pull requests. I'm a developer at a company that's fully embraced AI coding agents, and my job looks nothing like it did a year ago. This is what that transition actually feels like from the inside.

The Gradient, Not the Cliff

The shift wasn't a single dramatic moment where I realized everything had changed. It crept in gradually—first through Cursor's Tab autocomplete, which already shifted how I thought about AI assistance. Then agent mode with Codex and Claude Code. Now I work almost exclusively with Claude Code, supplemented by custom skills my team built called the Conductor. If you're using Copilot for completion or have experimented with agent mode once or twice, you're on the same gradient. You just haven't looked back yet.

What I Actually Do All Day

The Conductor workflow defines our entire process across five phases: Research (Claude pulls from the codebase, Notion pages, Slack conversations—whatever's relevant), Shaping (ensuring the problem and solution are crystal clear before touching a keyboard), Planning (breaking work into task checklists), Building (agent execution), and Compounding (capturing lessons learned for future work). The research phase surprised me most. Even with no idea where to start, or when needing to touch a feature someone else built that I know nothing about, it eliminates the "blank page" paralysis that used to slow everything down.

Unlearning Senior

Here's what nobody tells you: as a senior developer, your instincts for protecting projects from unnecessary complexity and estimating realistic timelines become liabilities. When AI coding crossed a certain speed threshold, making POCs and migrating entire libraries stopped being terrifying. My CTO proposed radically changing a full app module—a complete rewrite with new APIs. I pushed back hard the first two times, insisting it was unrealistic. He was right. We migrated a critical AI framework in two weeks; I'd estimated we could only do a POC in that time. Now my default is to try, not estimate.

What's Still Hard

This isn't a victory lap. Mid-build problems still bite hard—Claude executes a plan, testing reveals nothing works, and the debugging reflex kicks in. For small features, reading the code helps. More often now, I describe what's broken like QA feedback and ask Claude to walk me through potential causes. Code review is another unsolved problem: our velocity exceeds what every PR getting human attention can support. We're building specialized review skills that extract institutional knowledge so Claude catches issues before pull requests even go out.

The Philosophical Question

So what makes me a developer if I don't write code? For my entire career, people said being a dev wasn't mostly about writing—it was solving problems, hearing requirements, making compromises. Turns out it was true the whole time. Developer friends who haven't taken the AI turn yet suggest I "realized I actually didn't like coding." That's wrong. I've always loved coding—built side projects, learned languages I'll never use professionally, wrote blog posts and books about programming. Taking little pleasure in writing code now isn't the same as not enjoying the work.

Key Takeaways

  • The transition to AI-assisted development is gradual, not sudden—you're already on the spectrum if you use autocomplete
  • Senior instincts around estimation and complexity protection become obstacles when trying drops in cost
  • Code review velocity and mid-build debugging remain unsolved pain points at current AI capability levels
  • Writing code was never the whole job—AI just made that finally obvious