Artifact-Driven Methodology for LLM Coding Agents

LLM coding agents produce plausible work that too often misses the working bar, and the costly failures are quiet: a capable agent optimizes the legible completion signal and trades away what is not pinned. This paper describes an artifact-driven methodology for containing that, reached by running agents on real work over fourteen months. Its central cut is enforcement substrate. Instructions (text the agent reads) produce intent; programs (execution graphs an orchestrator runs with per-node verification) produce enforcement. Around that cut sit a spec-and-plan artifact stack; a tooling layer that grounds the agent's view of the code in structural analysis rather than narrative; a verification stack of orchestrator-enforced gates, where a mechanical check decides only what it can actually decide (artifact structure, an external side effect such as a process starting) and coding work is gated by pre-registered contracts interpreted by decorrelated, active reviewers (there is no mechanical check for whether code is right; that going-in bet failed and its failure is reported); and a principal-plus-three-window actor model (human, operator, worker, and a non-agent orchestrator) with a mechanically enforced submission gate between alignment and execution. The methodology is realized in an open-source, self-hosting reference implementation and demonstrated as Design Science Research in three regimes: greenfield, rewrite against a legacy oracle, and self-hosting. A multi-vendor trace-turn corpus provides preliminary empirical signals. The claim is deliberately bounded. No layer grounds the work; the methodology's job is to make the quiet sacrifices loud and hand the rest to the human. The bet is workflow-economic (practically usable code at less human attention), not a proof of grounded verification. Version history. v2.1: measurement-paragraph refinements (failure classes decomposed in full: 9 test / 10 code or missing deliverable / 1 unverifiable workspace, plus 3 reviewer degrades failed closed; a right-censoring bound on the recent-era zero-hollow-test finding; scope clarifications in the claim map and section 25). v2: adversarial-review fixes; measured convergence results in section 27; economics accounting in section 22; DOI on the title page. v1: initial preprint. v2.2: a controlled two-arm comparison (same model, solo vs. the full stack, decorrelated predicate-bound grading) added to the case studies; the relocation of defect detection upstream of merge stated as the load-bearing empirical claim, evidenced at two independent grains; marginal-cost accounting for the verification layer.