Specification-First Design of a Multi-Agent Short-Form Video Production Pipeline: Typed Stage Contracts, Offline Contract Validation, and an Audit of Evaluation Gaps

Abstract

Producing short-form vertical video well takes many coordinated roles. This report describes the Agentic Reel Studio (ARS), a specification-first design for a pipeline of large language model (LLM) agents covering that chain, from strategy through editing, sound, motion graphics, and quality control. Each agent owns one stage and exchanges typed JSON with its neighbors. Every agent passes through a gated lifecycle of dated artifacts: an outline, a specification body, a lock candidate with a reviewer checklist, Draft 2020-12 input and output schemas, a skill bundle with five fixtures, an offline validation harness, and an LLM test runbook. The public repository contains 13 skill bundles (65 fixtures), 9 schema pairs, 9 harnesses, and 11 runbooks. Re-executed for this report, the harnesses pass all 557 checks they report, and 273 of those checks confirm that mutated, contract-violating outputs are rejected. We separate what this evidence shows from what it does not. The harnesses verify contracts and hand-authored expected outputs; they do not exercise model behavior. None of the runbooks contains a recorded execution, so no measured LLM pass rates exist. Agents #5 to #10 were locked on offline evidence, and the v1.0 post-production agents emit specifications for human finishing rather than rendered media. We document failure modes addressed during design and latent fixture inconsistencies found by this audit. The contribution is a documented method and an accurate status record for a single-author, single-brand pilot, not a performance claim.