Spec-Driven Agentic Development (SDAD) is the name SDAD: Spec-Driven Agentic Development for the AI-Native SDLC gives to a software delivery methodology that combines disciplined up-front formalisation with high-velocity agentic implementation. Its authors argue that once coding agents can ingest entire repositories and requirement documents, the binding constraint on delivery becomes the clarity of intent rather than the speed of writing code, and that practice therefore converges toward the formal, pre-emptive specification associated with Waterfall — which they present as a maturation of practice rather than a regression.
Usage
The report defines SDAD as four phases. Intent capture uses structured human–agent dialogue to record requirements, constraints and business rules; formal specification turns that record into a machine-interpretable blueprint of interfaces, logic, data structures and acceptance criteria; agentic synthesis treats the specification as the authoritative input for multi-file generation of code and tests; and independent verification checks the output against the specification through automated test, security and conformance gates. Merge or release authority remains with an accountable human, whom the report calls the Spec Architect, rather than with the synthesis agents.
Around this workflow the report proposes an SDAD-V model patterned on the classical V-model; a redistribution of roles across engineering, QA, platform and product functions; metrics meant to replace story points, including Spec Fidelity, the Ambiguity Tax and a Synthesis Efficiency Ratio; hybrid estimation; and a gated migration path from Waterfall or Scrum. It frames SDAD as a mixed regime with human accountability, not a human-absent pipeline.
When It Applies
The report presents SDAD as applying where frontier agents with large context windows and multi-agent orchestration are available to synthesise cross-module features from a specification, and where a team can produce specifications precise enough to drive them; it assumes that specification quality, verification gates and provenance logging are made first-class. It names its own failure modes: a vague or malicious specification propagated across many files or services in a single synthesis pass, human comprehension of agent-authored code falling behind, and dependence on closed model vendors. Its standing is that of a single proposal: the report is largely positional, labels its cost figures as illustrative rather than audited, and lists the operational definition and measurement of its own metrics among its open questions.