Two Kinds of Engineers in 2026

There are two kinds of engineers in 2026.

The first kind says: “AI will do everything. Why should I learn anything?”

The second kind says: “AI can do a lot. What can I do with it now that I couldn’t do before?”

For a long time, I was the first kind. Both groups use the same tools, and they end up in very different places. Here is why I think the second group is ahead.

New tools create new problems

Every coding agent, RAG system, or AI harness that grows runs into problems that did not exist two years ago.

Take cost. The price gap between models is huge. One 2026 engineering guide puts the spread between the cheapest usable model and a frontier model at roughly 100x. That is why model routing exists: send easy requests to a small model and escalate only the hard ones. The RouteLLM paper reported up to 85% cost savings on the MT-Bench benchmark while keeping 95% of GPT-4’s quality. That is a benchmark result, not a guarantee for your traffic, which brings us to the second problem.

Routing to a cheaper model can quietly make answers worse, and nothing on your cost dashboard shows it. The usual fix is an evaluation gate: a set of 50 to 500 representative test cases that must pass before a routing change ships. Inference cost, latency, routing and evaluation are all jobs that someone has to do, and few people know how to do them well.

You only see those problems by building

If you decide AI will do everything, you stop building. If you stop building, you never run into the problems companies are hiring for.

Most of what I know came from things failing in ways I did not expect:

• Retrieval returning the wrong chunks. When a RAG answer is wrong, the instinct is to blame the model. Often the model did exactly what it was told, and retrieval handed it the wrong evidence. A policy document split mid-section can leave out the one paragraph with the actual dollar amounts, and the model answers honestly from incomplete context (a good walkthrough from AWS).
• An agent stuck in a loop. Agentic systems add failure modes that classic RAG does not have, such as retrieval thrash, tool-call storms, and context bloat (Towards Data Science).
• A task queue backing up. Plain old infrastructure, now with LLM calls that take seconds instead of milliseconds.

None of that would have happened if I had kept assuming the model would handle it.

The demand for help is already here

The hiring data points the same way. Lightcast’s Global AI Skills Outlook reports AI-skill job postings up 109% from 2024 to 2025. LinkedIn’s 2026 Jobs on the Rise ranked AI Engineer as the fastest-growing job title in the US. PwC’s 2025 analysis found a 56% wage premium for roles that need AI skills.

The more interesting number is about trust. In the Stack Overflow 2025 survey, 84% of developers said they use or plan to use AI tools, but only 29% trust the output, down from the year before. Hiring guides now say the strongest signal of real LLM experience is whether you can walk through an evaluation you designed.

That gap between “everyone uses it” and “few trust it” is where the work is. And you do not need a job title to step into it. Many people still pay for weaker tools, or use Claude Code at a fraction of what it can do. Teaching them, setting up their workflows, and building internal tools for them are services people pay for today.

What the belief costs you

If you believe AI will do all the work, and you also do not have enough money, the next question comes quickly: “What am I supposed to do with my days?”

That belief does not only cost you opportunities. It takes away your reason to get up and work on something.

What changed for me was going back to building. I now build LangGraph agents and RAG pipelines every day, and I learn something new each time one of them breaks.

The pessimist has a prediction

Maybe AI will do most of this work in 20 years. Nobody knows.

What you can control is the next few months: what you build, what you learn, and which problems you notice along the way.

The pessimist has a prediction. The optimist has a list of problems to work on.

Which problem have you hit while building with AI that nobody warned you about? Email me at me@ananthanarayanan.com. Even one line is fine, no need to overthink it. I read every email and will reply to each one.