SDAD: Spec-Driven Agentic Development for the AI-Native Software Development Life Cycle
Authors: Vu Hung Nguyen, Thanh Nguyen
Submitted: 5 May 2026
Subjects: Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
Abstract
Frontier coding agents—backed by large language models (LLMs) with context windows spanning hundreds of thousands to millions of tokens—are fundamentally reshaping the Software Development Life Cycle (SDLC). Their capacity to ingest large functional requirement documents (FRDs) and repository context in a single workflow, combined with multi-step reasoning, makes specification quality the primary driver of autonomous delivery performance. This paper formalizes Spec-Driven Agentic Development (SDAD) as a synthesis of disciplined, up-front formalization and high-velocity implementation. SDAD comprises four core stages: intent capture, machine-readable specification, agentic synthesis, and independent multi-agent verification under human sign-off.
We revisit the historical pendulum between Waterfall and Agile methodologies, introduce AI-code as a fourth production paradigm, and compare Human-Agile (circa 2020) with Agentic-SDAD (circa 2026) across key dimensions including artifacts, cadence, accountability, and security posture. Building on this process foundation, we extend the model to address team role metamorphosis—covering engineer, QA, platform, and product functions—and introduce quantitative governance metrics such as the Ambiguity Tax, Spec Fidelity, SER, and TCI_agentic with a repair multiplier (phi). We also propose pragmatic adoption strategies through hybrid estimation and a staged migration blueprint.
Integrating industrial and research evidence on AI-augmented testing and verification, we motivate the separation between synthesis and release authority. Overall, this paper argues that agentic speed does not eliminate engineering discipline—it relocates discipline upstream into specification precision, explicit gates, and auditable provenance, ensuring reliability and traceability in the AI-native SDLC.
via ArXiv AI