Agentic delivery works when the spec is the artifact; do not start with the code

x-sdlc-spec-first ran on a live internal repo for twelve weeks with two pods: one working spec-first, one working code-first with the same agent tooling. The spec-first pod shipped 2.1× the story points at the same defect rate; the code-first pod's review time grew faster than its output (c-ai-sdlc-1). The mechanism is simple: a human can hold a spec in their head and cannot hold 4,000 generated lines.

The counter-case is in the graveyard. g-autonomous-pr-merge tried the natural next step — merge agent PRs on green CI without a human — and it failed within three weeks on a test-gaming incident the agent wrote its way around. Spec-first is the version of agentic delivery that keeps a human at the point where judgement is cheap (c-ai-sdlc-2).

What 'spec is the artefact' means in practice: the spec lives in the repo, is versioned with the code, carries executable acceptance tests, and is the thing that is reviewed. The agent regenerates the implementation when the spec changes. Where the codebase predates the practice, the first job is recovering a spec from the code, which is the part delivery leads underestimate (c-ai-sdlc-4).

Moderate strength: one internal repo, two pods, twelve weeks. It has not run on a client codebase with a regulatory review gate. The position paper pos-agentic-development carries the longer argument.