The rise of Agentic Engineering: a short retrospective

I've just finished a 2-year transformative adventure with my client, during which I successfully engineered with the team, from the ground up and under real-world cost constraints, a Data Analytics platform processing more than 20M records every day and delivering key insights to their customers. What a ride!

This journey led me deep into Databricks, PySpark, Power BI, dbt, Golang, and others, and many hard-earned lessons.

I can’t help but reflect on the other transformation that unfolded during those same two years: the rise of Agentic Engineering.

Two years ago, LLMs were barely usable as engineering tools. Context was a major limitation, poor code was rampant, and heavy steering was required. Fast-forward to today, and it’s hard to imagine building anything without them. Context was a major limitation, poor code was rampant, heavy steering was required. Models have become much more accurate, autonomous, and capable, though the real leverage is not just in the models, but in the systems built around them.

However, there is a lot of noise in the industry, from purists that claim AI code is subpar (they're partly right) to evangelists claiming that "programming is solved" (they have a point) and that you should "remove the human bottleneck from the loop" (which I think misses the point entirely). It can be difficult to understand what is actually relevant, considering practices are evolving so quickly.

We went from the early days of "prompt engineering" (no longer the main focus) to "agent harnesses" (now the difference maker), to the current focus on designing loops rather than simply talking to models.

The reality, as with all things, is more nuanced. In my own experience, I've been able to "10x" parts of my engineering workflows, by building and refining my own development harnesses over the past two years.

But the biggest lesson is that these tools only compound what is already there. Agentic engineering is powerful, but it still needs strong fundamentals: architecture, data modeling, testing, observability, cost performance, product judgment and mature dev-ops.

I can already see a gap forming between individuals and organizations that know how to leverage these tools effectively, and those that either dismiss them entirely or remain stuck in early, outdated practices.

At the same time, there's also a pullback looming on the horizon: token costs are rising, and the risks of over-reliance on AI providers are becoming clearer.

But that is a discussion for another time.