AI coding agents are evolving from autocomplete tools into autonomous coding agents capable of understanding and implementing entire software specifications. This shift raises a key question: how do we ensure that what these agents build truly reflects what we intend? Spec-driven development offers an approach where natural-language or structured specifications become the central artifact that guides, tests, and validates code generation.

This talk explores the emerging practice of building software through coding agents that operate directly from specifications. We’ll provide an overview of the spec-driven ecosystem, from prompt structuring and planning to iterative testing and refinement. Using a real application built entirely via this method, we’ll walk through how the spec was written, how agents interpreted and implemented it, and how human feedback shaped the outcome.

We’ll share practical lessons on what works such as clear spec design, modular task decomposition, and self-validation loops, and what doesn’t, such as underspecified intent, context drift, and limitations. You'll see the impact of using the same specification for generating the app in different languages. Attendees will leave with a framework for incorporating spec-driven workflows into their own development processes, understanding both the potential and pitfalls of autonomous coding systems.

Jfokus

Arun Gupta presents spec-driven development (SDD) as a structured alternative to vibe coding when working with AI coding agents. The core idea is to invest heavily in writing detailed natural language specifications before generating any code, separating 'what' (the spec) from 'how' (the implementation). Key practices include: writing comprehensive specs with functional/non-functional requirements and Gherkin-format acceptance criteria; using agents.md to give coding agents persistent context and guardrails; creating reusable agent skills (portable engineering patterns like commit formats and subsection implementation commands); and generating implementation plans from specs before writing code. The talk uses a tic-tac-toe multi-agent application as a live example, featuring a 4,500-line spec and 600+ unit tests across six phases. Gupta argues that AI amplifies existing organizational strengths and weaknesses, so investing in specs and structure is essential. He references Andrej Karpathy's shift from 80% code/20% spec to 80% spec/20% code as validation of this direction.

Spec-driven Development using Coding Agents, by Arun Gupta