I just merged a pull request that Gemini estimated would take 4 to 6 weeks (20-27 person-days) of “exceptionally high skill” (Principal Engineer level) work.

It was a massive re-architecture of a core, stateful service. The final PR involved:

• ~4,000 lines for a new non-blocking scheduling strategy that maintains a budget of operations.
• ~8,300 lines for a complete test suite rewrite, including a new simulation harness.
• ~3,500 lines of new engineering standards and design documentation.

I delivered it in 3 days.

Here’s the most interesting part: I’m not sure I personally wrote a single line of code in that PR. And yet, it is very much my work.

This is a case study in the new role of the AI-augmented engineer.

The Challenge: The “Impossible” Bug

We were tackling a challenging problem in a complex, stateful system—the kind of scheduling and resource-budgeting bug that is notoriously difficult to reproduce, let alone fix. You can’t just run the service and “see” the bug; it only appears at scale, under specific, hard-to-predict race conditions.

The old way to solve this would have been weeks of logging, debugging, and writing brittle, implementation-specific tests, all while the production system continues to have issues.

The New Workflow: Engineer as Director

Instead of writing code, my role shifted entirely to that of an architect, director, and real-time reviewer. I was in a high-intensity “flow state” with my AI partner, Claude Code, over 123 commits.

My job wasn’t to type; it was to demand.

• I demanded architectural purity: I tasked the AI with refactoring the entire system, decomposing it into pure functional components and a small, imperative core.
• I demanded testability: I had the AI build an expressive simulation harness before we even fixed the bug. This allowed us to create the conditions of the bug in a local, deterministic test environment.
• I demanded quality (The “Guardrails”): I set up strict guardrails. I configured pyright for static type checking and wrote CLAUDE.md files with explicit instructions for the AI on how to handle errors and write code that meets our standards.
• I demanded iteration: As the simulation harness ran, I learned more about the problem. I had the AI redesign the scheduling and budgeting logic twice as my understanding improved—something that would have been cost-prohibitive to do manually.

My focus was on commit-level code review, reflecting on pitfalls, and continuously raising the bar for what I expected. The AI handled the generation, but I was the gatekeeper of quality and design.

The Result: Weeks to Days

The AI-driven workflow didn’t just 10x our speed; it fundamentally changed the quality and scope of the solution.

A human engineer (even a team) would have spent 3-5 days just designing the new test harness before writing a single line of code. We built the new architecture, the professional-grade simulation, and fixed the bug in that same timeframe. We found and fixed bugs in our simulation that would have taken months to pin down in production.

This is the power of this new partnership: it’s not just about writing code faster. It’s about having an AI partner that can instantly scaffold an entire high-quality architecture, a complete test suite, and all the documentation, allowing you—the human expert—to operate purely at the level of design and strategy.

You just have to know what to ask for.