At the beginning of 2026, it’s now clear to many that AI works — and works very well.
Experimentation is everywhere: everyone is testing, exploring, adjusting, in a continuous flourishing of ideas, workflows, and practices.
When exchanging feedback with my peers, the same refrain regularly comes up:
BMAD, Spec-Kit, or other “turnkey” frameworks supposedly designed to effectively structure work with AI.
“You want to drive development through specs? You should use Spec-Kit.”
And for the umpteenth time, we share a link to a GitHub repo, as if the answer were already there.
This reflex is not surprising.
In the world of “traditional” software, the framework is a rational response to a real problem.
Tools like Spring, Ruby on Rails, or React emerged to pool costly abstractions to design, difficult to maintain, and critical to avoid constantly reinventing fragile solutions.
Their implicit promise was simple:
inventing a wheel is difficult, time-consuming, and risky — and you’re likely to build a square wheel.
In most cases, this promise was kept.
Software history is marked by “homegrown” attempts that confirmed a simple truth: criticism is easy, art is difficult.
The error is not in the reasoning…
it’s in its mechanical application to a technology that radically changes the rules of the game.
With AI, the cost of creation — both technical and cognitive — has dropped dramatically.
Writing a prompt, structuring a context, defining an agent workflow or an MCP no longer has anything to do, in terms of effort, with designing a traditional software framework.
Today, a single person can, in a few hours:
In this context, adopting an existing framework also means importing:
The spectacular gains observed with AI don’t primarily come from tools.
They come from the quality of structured reasoning around AI.
In practice, the truly determining factors are:
No framework can decide this for you.
The best results observed in the field often come from simple, explicit workflows, sometimes even “cobbled together” — but perfectly aligned with the problem being addressed.
Rejecting frameworks wholesale would be a symmetrical error.
AI frameworks provide real value when:
In other words:
The problem is therefore not the existence of AI frameworks, but their premature adoption, often guided by hype rather than real need.
Conversely, creating your own AI framework is not free either.
Without explicit discipline:
A framework — even an imperfect one — sometimes forces one essential thing: making decisions visible and transmissible.
This is a cost that should not be ignored.
The real question is not:
“Which AI framework should I use?”
But rather:
“What do I really need today?”
“Am I exploring, or industrializing?”
“What constraints am I willing to accept — and which ones can I avoid?”
Creating software has never been easier.
Creating relevant AI workflows hasn’t either.
AI frameworks are neither miracle solutions nor mistakes to be avoided at all costs.
They are tools, useful at a given moment — but incapable of thinking for you.
And perhaps that’s the most important change brought by AI: the responsibility for structure now fully returns to humans.
Last modified: 2026-01-14