LLMs and Agility: Why Pit Them Against Each Other?
In recent weeks, the idea has been circulating that Agility is dead… because the cost of development has supposedly dropped to zero. In other words: there’s no longer any point in iterating, since everything can be built in one go.
Let’s just say it right off the bat: that’s absurd.
First of all, no, the cost of development hasn’t dropped to zero. Our work is accelerated, our capabilities are greatly amplified, but anyone with even a modicum of concern for the long term knows full well that the magic lies on the next level up. In fact, if you want something that works today—and only on the surface—then yes, you could say it will cost you almost 0—but that’s, at best, a prototype or a demonstrator. On the other hand, if you want to build an ambitious product designed to last for years, it’s in your best interest to put serious safeguards in place and pay close attention to the code generated, because that’s your technical foundation. If not handled properly, LLMs turn into veritable machines for creating technical debt.
But above all, this idea assumes that development cost is the only reason to proceed in iterations, which is false. Certainly, development cost is a reason to break things down into User Stories, but it’s neither the only nor the primary one. What motivated the founders of Agility to work this way was the desire to obtain frequent user feedback to ensure they were building in the right direction. In other words, to avoid getting stuck in a 6-month tunnel only to discover at the end that the developments do not meet user expectations, when a user test based on mockups could have provided the same information in a matter of days. That is what Agility is all about (among other things).
While AI is revolutionizing the way we work, one thing hasn’t changed: we still don’t know what’s going on in users’ minds. An LLM isn’t going to tell us whether we’re solving the right problem or working on the right feature, let alone whether there’s a market for it. So, while LLMs allow us to produce iterations more quickly and with better results, we still need just as many iterations as before.
In fact, LLMs are a double-edged sword:
-
On the one hand, they represent a real opportunity for Agility, because they allow you to get feedback more quickly. It’s also worth noting that fine-grained decomposition helps limit the scope of the work context, perhaps even more effectively than any other approach. Rather than pitting AI against Agility, we would therefore benefit from combining them.
-
At the same time, LLMs make the need for agility even more critical, because the faster we produce, the more the underlying complexity increases (it grows with the number of features, even if the code is of very high quality). It is therefore more important than ever to ensure that these features truly provide value to users; otherwise, they only cause inconvenience.
Agility isn’t dead—quite the opposite. As for LLMs, they mainly help us with delivery but also with discovery. So, why pit them against each other?