With the rise of LLMs, a new promise has emerged: software development driven by specifications. Multi-agent systems, RAG, MCP... everything seems to converge toward an appealing idea: given enough context, AI can build an entire system.

In practice, however, this approach—often referred to as Spec-Driven Development (SDD)—quickly reveals its limitations. It assumes a perfectly understood context, known invariants, and an AI capable of filling every gap in the specification. The result is often the same: systems that look impressive in demos but fail in production, leaving behind a massive understanding debt.

This talk shares concrete lessons learned from real-world projects where AI is used daily in complex product environments.

We will explore:

Why SDD is a modern reincarnation of the Waterfall model, fundamentally incompatible with the complexity of real-world systems.
How human cognitive limitations (bounded rationality) also apply to the way we use LLMs.
Why treating an LLM as an unreliable accelerator radically changes the way we build software.

We will introduce a pragmatic alternative: Empirical Product Development (EPD), an approach built around:

Atomic tasks
A wall of feedback (tests, performance, system consistency, business validation)
Iterative and controlled progress

Through real examples, we will show how to:

Integrate AI into an agile development cycle without losing control
Avoid black-box effects
Accelerate delivery without compromising quality

This talk is aimed at practitioners already using AI in their day-to-day work who want to move beyond the hype and build systems that are truly robust in production.