AI Is Reshaping Developer Work—But Not the Way You Think

You've probably heard that AI will change your job. What you might not know is that it's already changing it in unexpected ways—and the shift isn't about replacement. It's about repositioning.

Three converging signals reveal what's really happening: new research challenges popular AI coding practices, industry conferences focus on making AI tools actually usable, and the skills gap is widening in a direction you might not expect.

The Context File Paradox

ETH Zurich researchers just published findings that challenge a practice thousands of developers have adopted: adding AGENTS.md files to their repositories. According to InfoQ, over 60,000 open-source repositories now contain these context files, meant to guide AI coding agents through codebases.

The research tested four AI agents across 138 real-world Python tasks. The results? LLM-generated context files reduced task success rates by 3% and increased inference costs by over 20%. Even human-written context files, while marginally better with a 4% success boost, drove costs up by 19%.

Why? The AI agents followed the instructions dutifully—running more tests, reading more files, performing more checks. They appeared thorough. But they were often solving the wrong problem, spending tokens on work that didn't matter for the specific task at hand.

Here's the real lesson: AI tools excel at execution but struggle with judgment about what matters. The researchers found that context files didn't help agents locate relevant files or understand architecture. They just made agents do more work.

What DeveloperWeek Revealed About Usability

At DeveloperWeek 2026, the conversation centered on a question developers are asking more often: are AI tools actually making me faster?

Caren Cioffi from Agenda Hero shared a relatable frustration in her session: trying to get an AI image generator to create the perfect image. The first attempt was almost right. Each subsequent attempt to fix small details made things worse. Why? Because AI tools are black boxes that only accept natural language prompts.

According to the Stack Overflow coverage of the event, the usability problem is fundamental. Most AI tools are built for efficiency and speed, not for how easy they are for humans to use. As Cioffi pointed out, users need agency—the ability to edit small sections of AI output rather than regenerating everything from scratch.

The Stack Overflow keynote drove this home: context emerged as the buzzword of the conference. As Akamai's Senior Director of Developer Relations Lena Hall put it: "Context is all you need."

But here's what matters for your work: AI tools without company context generate code that doesn't match your architecture or standards. This leaves developers doing janitorial work—cleaning up and reorganizing output that was almost right but not quite.

Sound familiar? It's the same problem as those AGENTS.md files.

The Real Shift: From Solving to Judging

A post on DEV Community cuts through the hype with a simple observation: coding itself provides no value. It never did.

Businesses use computers to solve problems. Code is just the mechanism. In the past, developers spent time copying from Stack Overflow, building reference projects, and solving implementation puzzles. AI now handles much of that retrieval and assembly.

But as the author notes, when someone needs specific business value, they still need to hire someone who understands the problem. "If someone needs to hire someone, he will. If he can vibe code it, totally fine."

The bar is rising. A developer who only knows for-loops will need to understand while-loops when the AI uses them. Knowledge that was slowly accumulated over years now needs to be learned faster. As one 25-year veteran mentioned in the post: they haven't written much code in the last year because "they saw it all."

What This Means for Your Career

The evidence points to a clear inflection point. AI tools are good at implementation. They're not good at:

This creates a specific opportunity. As AI handles more implementation, your value shifts to:

Setting Better Context

Not AGENTS.md files that make AI do busywork. Real context about business constraints, architectural decisions, and domain knowledge the AI can't infer.

One developer commenting on the ETH Zurich research nailed it: "The biggest use case for AGENTS.md files is domain knowledge that the model is not aware of and cannot instantly infer from the project. That is gained slowly over time from seeing the agents struggle due to this deficiency."

This is your competitive advantage. You accumulate this knowledge. AI doesn't.

Making Architectural Decisions

AI can generate components. It can't decide your infrastructure or make judgment calls about technical debt. Those decisions require understanding tradeoffs that extend beyond the current task.

Validating and Editing Output

Developers are already spending more time reviewing AI-generated code than writing from scratch. But effective review requires knowing what good looks like—understanding patterns, recognizing anti-patterns, and catching subtle bugs.

Communicating with Stakeholders

AI can't negotiate requirements, push back on unrealistic timelines, or explain technical constraints to non-technical teammates. These skills become more valuable as the technical implementation becomes easier.

What to Do Today

Stop treating AI tools as magic boxes. Start treating them as junior developers who execute well but need direction.

Document your reasoning, not just your code. When you make an architectural decision, write down why. This creates the domain knowledge AI can't access. It also helps your team and your future self.

Focus on the problem space, not the solution space. Spend more time understanding business needs and less time optimizing implementations AI can handle.

Learn to set better context. Experiment with what information actually helps AI tools produce better output. The ETH Zurich research suggests it's not architectural overviews—it's specific, non-inferable details about tooling and build processes.

Practice editing, not just prompting. Get comfortable quickly iterating on AI output by directly modifying it. The most productive developers won't be the best prompters—they'll be the ones who can efficiently guide AI toward the right solution through a combination of prompts and edits.

The fundamental change isn't that AI is replacing coding. It's that coding was never the point. The point was always solving business problems with computers. AI is just making it clearer where the real value lies.

And that value has always been in your judgment, your context, and your ability to understand what problem actually needs solving.