source: arXiv Artificial Intelligence: SDAD: Spec-Driven Agentic Development for the AI-Native SDLC
level: research
a new arxiv report introduces spec-driven agentic development, or sdad, as a formal method for building software with ai coding agents. the approach combines upfront specification with fast implementation. it has four parts: capturing intent, creating a machine-readable spec, letting agents write code, and using multiple agents to verify the work before a human signs off. the authors argue that as large language models handle more context, the quality of the specification becomes the main driver of successful autonomous delivery.
the report compares human-agile methods from around 2020 with agentic-sdad from around 2026. it looks at differences in artifacts, development cadence, accountability, and security. the authors note that frontier coding agents now have context windows from hundreds of thousands to millions of tokens. this lets them ingest large functional requirement documents and repository context in one workflow. however, the report does not provide empirical benchmarks or case studies, so the claims remain conceptual rather than proven with data.
the paper revisits the historical shift between waterfall and agile methods. it positions ai-generated code as a fourth production paradigm, after hand-coded, model-driven, and agile approaches. the authors suggest that sdad may reduce manual coding effort but increase the need for precise specification skills. for data science teams, this could mean more time spent on defining requirements and less on implementation. the report is a preprint and has not yet undergone peer review.
why it matters: for ai and data science teams, sdad could shift effort from writing code to writing precise machine-readable specs, changing how projects are planned and verified.
source: arXiv Artificial Intelligence: SDAD: Spec-Driven Agentic Development for the AI-Native SDLC