Deterministic Data Engineering With AI Harnesses: Using Claude Code, Codex, Antigravity, and OpenCode for Data Work You Can Actually Trust

This article explores the apparent contradiction between probabilistic AI agents and deterministic data engineering. It presents an architectural principle: use AI agents to author deterministic artifacts (code) that execute reliably, rather than letting agents improvise at runtime. The author introduces a 'determinism ladder' with five levels of trust, profiles four AI harnesses (Claude Code, Codex, Antigravity, OpenCode) as data tools, and provides a catalog of techniques (tests as contracts, dry-run gates, schema pinning, semantic layers, golden datasets) for reproducible data work. Workflow patterns for pipelines, migrations, quality investigations, and analytics are covered, along with anti-patterns that cause incidents. The article is a practical guide for data teams seeking productivity gains from AI without sacrificing data integrity.