Every year, the industry announces a new “game-changing” innovation. Most of them matter far less in practice than they do in talks and blog posts. A few are useful. Very few actually change how software is built.

AI agents did.

In 2025, AI stopped being an experiment and became infrastructure for how I build software. It now shapes how I think about problems, how I design systems, how I write code, and how I review the work of others. Ignoring it today feels similar to ignoring version control or automated testing once they became standard.

When the Conversation Changed

The shift didn’t happen all at once.

I first noticed it in July at We Are Developers 2025 in Berlin. AI was clearly a trend there—present in many talks, widely discussed, but still framed as something emerging.

Two months later, at GOTO; Conf 2025 in Copenhagen, the tone had changed completely. AI was no longer a trend. It was a reality. And not just any conference—GOTO consistently attracts speakers who have been shaping this industry for decades. When that crowd converges on the same conclusion, it’s usually worth paying attention.

The question wasn’t whether AI mattered, but how to use it without creating long-term damage.

At Turo, our engineering organization made an early and explicit bet on AI to increase engineer leverage. Not to replace people. Not to lower standards. But to turn strong engineers into force multipliers.

AI doesn’t replace engineers. It reflects the quality of the thinking behind it.

How AI Is Already Changing How I Build Software

The Tools I Actually Use

My setup is simple: Claude, Cursor, and ChatGPT. The value isn’t in the tools themselves, but in how deeply they’re embedded in the workflow.

I use them daily in real production work at Turo. This isn’t about side projects or demos. It’s about shipping code, reviewing pull requests, and making architectural decisions under real constraints. Organizational support made the difference between experimentation and real leverage.

These tools don’t think for me. They think with me.

My Three Core Modes of AI Usage

Over time, my usage settled into three clear modes.

1. Asking questions at machine speed
Architecture tradeoffs, framework behavior, edge cases, second opinions. AI collapses hours of context-switching into minutes.

2. Offloading the obvious
Boilerplate, glue code, repetitive transformations, mechanical refactors. AI handles the predictable work so I can focus on decisions that actually matter.

3. Pairing on hard problems
For complex problems, I don’t ask for answers. I debate. We design. We break things down. We implement step by step. I review, adjust, and commit incrementally. Done right, this feels like compressed collaboration.

A Concrete Use Case: Learning a Codebase Faster (Without Breaking It)

One use case I’ve found particularly valuable is using AI to understand the implicit conventions of a codebase I haven’t worked with extensively.

Instead of guessing style, structure, or patterns, I ask the agent to analyze existing code and explain:

• How responsibilities are typically split
• What conventions are followed (and which ones actually matter)
• How similar problems have been solved elsewhere

This allows me to produce code that fits—not just functionally, but stylistically and architecturally. The result is fewer review cycles, less friction with the team, and changes that feel native to the codebase rather than bolted on.

This is not about delegating understanding. It’s about accelerating it.

Why AI Makes Skill More Important, Not Less

AI Is an Amplifier

AI doesn’t flatten skill differences. It exaggerates them.

Strong engineers get faster and more precise. Weak foundations lead to faster accumulation of technical debt, overconfidence, and fragile systems that look correct until they fail.

A simple rule emerged for me:

AI increases the slope of your trajectory — it doesn’t change its direction.

This echoes ideas I explored years ago in What Makes a Great Software Engineer, long before AI entered the picture. The core qualities haven’t changed — evergreen skills still matter. More than ever. What’s changed is how quickly their absence becomes visible.

With the right context and constraints, AI feels less like autocomplete and more like leverage.

Vibe Coding vs Prompt Engineering

This distinction matters.

Vibe coding is fast, intuitive, and chaotic. It’s great for prototypes and MVPs.

Prompt engineering is deliberate and constrained. It’s how you build systems you expect to maintain.

Vibe coding gets you a demo.
Prompt engineering gets you a system you can live with.

This is also why I’ve been skeptical of simplistic productivity narratives for years. In On Software Engineers, Productivity, and Misleading Metrics, I argued that output alone is a poor proxy for impact — a problem that becomes even more visible when AI enters the picture.

Prompt Engineering Is Just Engineering

Good prompt engineering looks like good engineering: clear goals, explicit constraints, decomposition, validation. The interface changed. The fundamentals didn’t.

Context Makes a Huge Difference

This is also a key aspect. Without context, AI is mediocre. With context, it becomes genuinely powerful.

In practice, the biggest gains I’ve seen come from integrations that eliminate friction and surface the right information at the right time.

MCPs Are Key

Connecting AI to Jira, Confluence, New Relic, and GitHub allows agents to reason about real systems instead of abstractions.

Most “AI doesn’t work for us” complaints disappear once this level of context is in place.

Predefined Agents Beat Generic Ones

We’ve also had strong results creating purpose-built agents, each with a predefined role and context:

• An RFC-writing agent
• A product-manager-style agent focused on problem framing
• A senior software engineer agent biased toward best practices, maintainability, and risk assessment

These agents aren’t smarter models. They’re better situated. That turns out to matter more.

AI Code Reviews: Free Value, Today

What AI Reviewers Already Do Well

AI reviewers are excellent at the boring but important things: obvious bugs, inconsistencies, missing tests, and common edge cases. They don’t get tired. They don’t rush.

Raising the Bar for Humans

At Turo, AI acts as a first-pass reviewer. Humans spend more time on architecture, intent, and risk. The result is higher standards, not lower ones.

The Rise of the Technical Product Manager

This shift also connects closely to how I think about software as a discipline. Years ago, I wrote about the difference between building projects and building products. What’s changing now is that AI makes it increasingly possible for a single engineer to operate with a product mindset end-to-end — from problem framing to delivery — without needing a large team tackling separate “projects” to bridge the gaps.

One Person, Small-Team Output

This is the first time in my career where I can realistically see one engineer doing work that previously required a small, well-coordinated team. With AI agents, deeply technical product-minded engineers can define problems, design solutions, and execute end to end.

Programming Background Is Not Optional

This only works if you actually understand systems. Without that, AI accelerates the wrong things. With it, the leverage is enormous.

A Personal Direction

I could see myself moving into this kind of role. Not away from engineering, but deeper into it—closer to product and outcomes, with AI doing the mechanical heavy lifting.

The Dark Side

Fast Output, Slow Pain

AI makes it easy to move fast in the wrong direction. The short-term productivity gains are real. The long-term costs are too, if judgment is missing.

Experience Still Wins

Experience shows up in what you reject, not what you accept. AI doesn’t remove responsibility. It amplifies it.

Conclusion: The Bar Is Going Up

AI isn’t replacing software engineers. It’s changing which ones matter.

For engineers with strong foundations and product sense, this is an unprecedented leverage moment. For everyone else, it’s a trap.

The future of software development belongs to those who can combine judgment, fundamentals, context, and intent with AI at scale.

TL;DR: AI doesn’t replace engineers; it amplifies skill, judgment, and context — or the lack of them.

Oh, and yes — this article was mostly prompt-engineered.