How to build a decent product with Agentic AIs

For years, building software meant planning and coordinating the delivery of a complex set of tasks across a variety of disciplines, like research, design, development, quality assurance and, let’s not forget, project management. This often involved dozens of people with different levels of seniority and domain expertise, all of which needed to understand exactly what had to be done, in what order and by when.

This process was refined over years and years, across countless projects and industries. There are tons of certifications, methodologies and standards designed to make software delivery faster, more reliable and easier to coordinate.

Even most of the programming languages and technologies we use today exist solely because people needed faster ways to build software for scenarios that never stopped growing in complexity. Over the years, we created a kitchen sink of platforms, frameworks and tools to abstract away that complexity and make it easier to develop, debug and maintain solutions that, not too long ago, seemed impossible.

And I haven’t even gotten to the part where these solutions, their challenges and technicalities had to be communicated to business stakeholders who neither understood, nor cared that much about the details.

Now, the reason why I’m telling you all this, dear reader, is that all these resources, processes and disciplines play a very important role in the way LLMs work today. OpenAI, Anthropic, xAI and the rest of the large language model providers did not invent new technologies or processes. They pretty much took what was already there, strapped a rocket to it and put it under your fingertips for just a little over 20 bucks a month.

Today, you can use these LLMs to apply decades of accumulated software knowledge, practices and conventions to build pretty much whatever you want, given that you have enough tokens to throw at the damn thing.

Like all magical things, LLMs have a couple of faults. They are probabilistic, they lose context, they lack genuine accountability and ownership, and they hallucinate at the worst possible moments. So what’s the big deal? We’ve all worked with a couple of people like that, amirite?

‍