A talk from an AI engineer at Watershed describing lessons learned deploying coding agents to help users edit complex supply chain graphs for sustainability carbon footprint calculations. The core challenge: in a domain full of expert judgment calls, validating the final answer alone is insufficient — you must also validate the process. The solution involves constraining agent effects (not expression) via a typed TypeScript SDK as the only permitted interface for graph edits, combined with deterministic execution orchestrated by the harness. This catches false reports, prevents invalid actions, and produces human-readable review artifacts that non-coders can verify. Using this approach alongside prompt engineering and few-shot examples, eval accuracy improved from 43% to 92%.