SDAD formalizes spec-driven development for AI coding agents
Frontier coding agents built on large language models with context windows ranging from hundreds of thousands to millions of tokens are reshaping the software development life cycle (SDLC), according to a new arXiv report. Because these agents can now ingest full Functional Requirement Documents (FRDs) and an entire repository's context in a single workflow, the report argues that specification quality has become the execution fuel for autonomous delivery: how precisely a project's intent is written down now determines how well an agent can build it.
The report's core contribution is formalizing Spec-Driven Agentic Development (SDAD): a synthesis of disciplined up-front formalization and high-velocity implementation, built around four stages, intent capture, machine-readable specification, agentic synthesis, and independent multi-agent verification carried out under human sign-off. To place SDAD historically, the authors revisit the pendulum swing between Waterfall and Agile software development and introduce AI-code as a fourth production paradigm. They then compare Human-Agile, dated to around 2020, against Agentic-SDAD, dated to around 2026, across artefacts produced, delivery cadence, accountability, and security posture.
Beyond describing the process, the report extends its model to how team roles change under SDAD, covering engineer, QA, platform, and product functions, and introduces a set of quantitative governance metrics for tracking autonomous delivery teams: an Ambiguity Tax, Spec Fidelity, SER, and TCI_agentic, the last modified by what the report calls a repair multiplier, phi. The abstract names these metrics without publishing numeric values, formulas, or worked examples for them, and does not spell out what the acronyms SER and TCI_agentic stand for. For adoption, it proposes a hybrid estimation approach and a staged migration blueprint, though the abstract gives no concrete timeline for how long that migration should take.
To support the argument, the report draws on existing industrial and research evidence about AI-augmented testing and verification, using it to justify keeping synthesis, the work an agent does, separate from release authority, the decision to ship it. Its stated conclusion: agentic speed does not eliminate engineering discipline; it relocates discipline upstream into specification precision, explicit gates, and auditable provenance.
Key facts
- The report formalizes Spec-Driven Agentic Development (SDAD), a four-stage process of intent capture, machine-readable specification, agentic synthesis, and independent multi-agent verification carried out under human sign-off.
- It frames the large context windows of frontier coding agents, from hundreds of thousands to millions of tokens, as what lets a full specification and repository context be ingested in one workflow, making specification quality the execution fuel for autonomous delivery.
- The report introduces AI-code as a fourth SDLC production paradigm alongside Waterfall and Agile, and compares Human-Agile (circa 2020) with Agentic-SDAD (circa 2026) across artefacts, cadence, accountability, and security posture.
- It proposes four governance metrics for autonomous delivery teams, an Ambiguity Tax, Spec Fidelity, SER, and TCI_agentic with a repair multiplier called phi, though the abstract gives only their names, not numeric definitions.
- Its stated conclusion is that agentic speed does not eliminate engineering discipline; it relocates discipline upstream into specification precision, explicit gates, and auditable provenance.
Why it matters
As coding agents get large enough context windows to swallow a full specification and a whole repository in one pass, execution speed stops being the bottleneck for AI-driven delivery, the quality of the specification becomes the bottleneck instead. SDAD is an attempt to name that shift and give it a repeatable structure, positioning itself as a successor paradigm to Waterfall and Agile rather than a variant of either. Naming AI-code as a fourth production paradigm is itself a claim worth tracking: it signals that at least one report sees agentic coding as a break with prior methodology, not just an accelerated version of it.
Who it affects
The report is aimed at organizations already running or planning agentic coding at scale, and at the people inside them whose jobs it says will change shape: engineer, QA, platform, and product roles are all named as functions undergoing what the report calls role metamorphosis under SDAD. It is also relevant to whoever owns governance and release decisions for an autonomous delivery team, since the report's central distinction is between agents that synthesize code and humans who retain release authority.
How to use it
SDAD is structured as four stages: intent capture, turning a project's goals into a machine-readable specification, agentic synthesis of the implementation from that specification, and independent multi-agent verification that still runs under human sign-off before release. The report pairs this with a hybrid estimation approach and a staged migration blueprint for teams moving from a Human-Agile process toward SDAD, though the abstract does not give a concrete timeline for that migration. It also proposes four governance metrics, an Ambiguity Tax, Spec Fidelity, SER, and TCI_agentic with a repair multiplier called phi, for teams that want to track how well their specifications and agent output are holding up, though the abstract gives their names without the numeric definitions needed to compute them.
How solid is it
The report is a preprint on arXiv; it does not state a publication venue, a conference, or a peer-review status, and it names its authors, Vu Hung Nguyen and Thanh Nguyen, but not an affiliated institution, so its provenance and authority cannot be checked from the piece itself. It says it integrates industrial and research evidence on AI-augmented testing and verification to motivate its argument for separating synthesis from release authority, which suggests grounding in existing practice, but the abstract does not itemize that evidence. The four governance metrics it proposes are introduced as named concepts, not as validated instruments: no formulas, thresholds, or worked examples accompany Ambiguity Tax, Spec Fidelity, SER, or TCI_agentic in the abstract, and the acronyms SER and TCI_agentic are not spelled out there.
Risks and caveats
Because its authors are named but not their institution, readers have no way to weigh the report's expertise or any conflicts of interest from the text alone. No specific companies, products, or tools are named in the abstract as examples of the 'frontier coding agents' it describes, so it is not clear which real systems, if any, the report tested its framework against. The governance metrics it proposes are vocabulary without measurement yet: without numeric definitions in the abstract, teams cannot compute an Ambiguity Tax or a TCI_agentic score today. And the 'fourth paradigm' framing, like the staged migration blueprint's timeline missing from the abstract, is the report's own proposal rather than an outcome that has been independently verified.
“agentic speed does not eliminate engineering discipline; it relocates discipline upstream into specification precision, explicit gates, and auditable provenance”
— the paper