I Changed My Mind: AI Will Replace Us

For many years, I’ve been that engineer. You know the one: confidently explaining why our jobs are safe, why engineering is “too complex” for automation, why AI is just fancy autocomplete.

Well, I was wrong.

Looking back, over the past few years I can count four “oh shit” moments relating to AI and how it impacted my work (and life).

GitHub Copilot: “Super neat autocomplete,” I thought. Saved a bunch of typing. Still needed my brain for the hard stuff.

Cursor: The IDE became my pair programming partner. Not just completing lines, but understanding intent, refactoring entire functions, debugging complex logic. Cursor was also my introduction to “agentic” workflows.

Deep Research: Wait, this thing can actually research? Not just regurgitate, but dig, synthesize, connect dots I hadn’t seen. Saved a ton of time I usually spend researching various options for work or for personal stuff.

Lastly, a fresh experience from the past few weeks that triggered this post:

Roo Code + Claude Opus 4 + MCP: This is where I lost sleep. It’s a combination of tools that work wonderfully together and represent my AI-less workflow eerily similarly.

You could argue that these are not the deep-tech AI innovations that make a difference, but the truth is that it was these tools and not a completions API endpoint that had such an impact on my day-to-day that now I can see the future where humans won’t need to type a word to build software.

I dismissed Model Context Protocol (MCP) servers at first. “Another API wrapper,” I thought, while watching everyone rush to build MCP connectors for their services.

Then I actually used one. Then I used many.

Watching the developer agents look up terms in a data dictionary, metric definitions in a semantic layer, introspect my entire data warehouse through MCP, understand the schema relationships, identify inconsistencies, and propose fixes (all in minutes) was genuinely unsettling.

It wasn’t just reading documentation. It was understanding the business logic embedded in my data models.

For years, we data engineers have hidden behind the “business context” shield. “AI will never understand why we model revenue this way, or why customer churn is calculated differently for enterprise vs. SMB.”

Turns out, you can just… tell them.

In plain English. In comments. In documentation. Through MCP servers that expose your entire data lineage.

The AI doesn’t need to intuit business context, it just needs access to it. And now it has that access.

Yesterday, I watched an agent spend five minutes exploring my data warehouse, then propose a complete data quality framework with implementation details. It identified edge cases I’d missed, suggested optimizations I hadn’t considered, and generated dbt models that would have taken me days to write.

It was continuously verifying its own theories by being able to actually execute queries in the data warehouse.

I felt like a supervisor watching a competent engineer who never gets tired, never makes typos, and has perfect recall of every data engineering best practice ever written.

The barrier to entry is still real: you need to learn Cursor, set up MCPs, understand agents, manage context windows. It takes time and effort.

But that barrier is shrinking fast.

Engineers who invest in these tools now have a 1000x advantage. We’re like the first wave of developers who learned to use search engines and Stack Overflow effectively.

The main difference is that this time, the tools are getting exponentially better every month.

If AI can:

• Understand business requirements from documentation
• Explore and comprehend complex data schemas
• Write production-quality code
• Debug and optimize existing systems
• Work 24/7 without breaks

Then what exactly are we protecting?

Most of my engineering work involves data systems scattered across different departments. The biggest challenge is connecting these silos, centralization has become critical for keeping context close to where decisions are made.

“Roo Code + Claude Opus 4 + MCP servers” doesn’t feel like just a tool stack, but rather a preview of how data work gets done when AI can access everything, understand everything, and implement everything. In a few years (months?) you won’t need to combine so many different tools together, you’ll just have one thing to do all of these, probably all running in your smart glasses while you’re chilling on the beach.

The future isn’t AI-assisted development. It’s AI development with human oversight at most.

And that future is already here.