Spec-driven development is the industry's structured answer to vibe coding: write the intent down — constitution, spec, plan, tasks — and let the agent implement against it. By mid-2026 it stopped being an experiment. GitHub's spec-kit shows roughly 123k stars and templates for more than thirty agents github.com/spec-kit, Jul 2026 live figure, not adoption, AWS shipped Kiro as its mainstream bet, and the Thoughtworks Radar treats SDD as an established technique — a "feedforward control" for agents Radar Vol 34, Apr 2026. Adoption settled one question. The live one: is it agile?
§1The three rungs
The vocabulary that made the debate tractable came from Birgitta Böckeler's taxonomy Böckeler, martinfowler.com, Oct 2025: spec-first — the spec drives the task and may not outlive the feature; spec-anchored — the spec is kept and evolved as the maintained description of behavior; spec-as-source — the spec is the artifact you maintain and code is compiled output, stamped "DO NOT EDIT".
Her key finding survived into mid-2026: most tools and most teams are still only spec-first. Spec evolution over a feature's lifetime remains the unsolved problem. Compatibility with agile practice reads directly off the ladder — spec-first with small slices is unambiguously compatible; spec-as-source is the rung where "documents before code" genuinely bites, and it stays largely aspirational.
§2The waterfall accusation
The most damaging attack is empirical. Marmelab's François Zaninotto showed a single "display the current date" feature ballooning to roughly eight files and 1,300 lines of specification under Spec Kit Marmelab, Nov 2025 — heavy documentation before coding, revived. Kent Beck's critique circulated through Fowler's site soon after Fowler's Fragments, Jan 2026. The rebuttals concede the failure mode and relocate the difference: waterfall resists changing the spec once implementation begins; SDD done well rewrites it freely.
The verdict this collection stands behind: SDD is iterative waterfall in the small, wrapped in agile in the large — spec, plan, and implement phases at story granularity, hours to days, inside an agile outer loop. Keep the batch small and the spec editable and the accusation dissolves. Freeze an epic-sized spec and you have rebuilt 2003 with better autocomplete.
§3Drift, the highest friction
The mechanical weak point is reverse sync: code gets edited directly, and the spec no
longer describes reality. The 2026 tooling matured from "nothing" to "manual detection" —
spec-kit's /speckit.converge re-assesses the codebase against the spec and
appends the remaining work; OpenSpec and community extensions push toward bidirectional
sync — but every one of them is manual-trigger and human-reviewed
spec-kit repo; dev.to SDD comparison, 2026. Nothing self-heals.
Kiro exhibits the same drift between its generated requirements and evolving code
TechTimes, Jun 2026. The practical rule is unchanged: treat
drift as technical debt, reconcile deliberately, and do not trust a spec to track reality
on its own.
§4The constitution, standardized
One SDD idea won so cleanly it stopped being an SDD idea. The constitution — team standards as machine-enforceable text — converged on AGENTS.md, an open format transferred to the Linux Foundation and read by thirty-plus tools across tens of thousands of repositories AGENTS.md / Linux Foundation, 2026. Architectural constraints, operational mandates like test-first, style rules, negative constraints — injected into every agent session. Culture maintained through code turns out to be more robust than culture maintained through habit, and it is the cheapest piece of the whole stack to adopt. It is also the one piece the field agrees on.
§5What the data supports
The honest empirical footing is thin. The DORA 2025 report — about 5,000 respondents — found roughly 90% of developers using AI, with AI now correlating positively with throughput but still negatively with delivery stability DORA 2025. Its seven amplifying capabilities include working in small batches and strong version control — exactly the controls that keep SDD out of the waterfall ditch. But DORA measures AI-assisted development generally, not SDD; no controlled study isolating SDD's effect on delivery outcomes existed as of July 2026. The "3–10× first-pass success" figures that circulate are early-adopter reports vendor claim · no methodology. Adoption is verified. Value is not.
§6Practical guidance
What survives contact with all of the above: adopt the constitution pattern immediately — low cost, standardized, high value. Default to spec-first with small vertical slices; climb to spec-anchored only where a spec genuinely needs to outlive its feature; treat spec-as-source as unproven. Use SDD where complexity, risk, or compliance justify the ceremony, and exempt what fits it worst — UI polish resists text specification; hybrid with the loose end of the spectrum there. On brownfield systems, spec the delta, not the legacy; large codebases exceeding context windows remain SDD's hardest case. And keep the downstream review: DORA's stability finding says the plan review does not replace it. At the top of the ladder sits the maximalist position — essay 05 takes it apart.
§7Sources
- Birgitta Böckeler — Understanding SDD: Kiro, Spec-Kit, Tessl (Oct 15, 2025).
- Marmelab — The Waterfall Strikes Back (Nov 12, 2025); Kent Beck via Fowler's Fragments (Jan 8, 2026).
- github/spec-kit (accessed Jul 23, 2026); drift tooling — SDD framework comparison (2026); Kiro drift — TechTimes (Jun 17, 2026).
- AGENTS.md — format guide (2026).
- DORA 2025 State of AI-assisted Software Development; Thoughtworks — Radar Vol 34 (Apr 15, 2026).