Descriptions:
Andrew Dumit, an AI engineer at Watershed — the enterprise sustainability platform — delivers a conference talk detailing how his team evolved from a React agent to a coding agent architecture for a complex, expert-judgment-heavy task: helping users edit large supply chain graphs representing the full carbon footprint of manufactured products. Each graph can contain thousands of nodes with rich metadata, and a single product like dark wash jeans requires modeling upstream materials, transportation, energy, and industrial processes as a connected graph.
The core problem Dumit confronts is that in sustainability, correctness of process matters as much as correctness of output — six domain experts given identical wine production data produced answers varying by up to 50%, each defensible by their own methodology. This makes pure output validation insufficient. The solution Watershed landed on is what Dumit calls “constraining the effects, not the expression”: the coding agent is free to write arbitrary code, but all graph-editing operations must pass through a typed TypeScript SDK that enforces which fields are editable versus derived, validates schema compliance, and enables deterministic, traceable, replayable execution. The system can reject and retry if the agent’s code doesn’t produce valid typed objects.
Dumit shares the failure modes encountered along the way — including hallucinated schema fields as context windows filled up, and inconsistent approaches across multi-graph batches with the original React agent. The talk draws on a 2026 Open Proof Corpus paper to ground the process-vs-answer distinction in broader research. Engineers building agents for domains with expert-judgment variability or compliance requirements will find the typed SDK pattern and deterministic execution layer directly applicable.
📺 Source: AI Engineer · Published July 07, 2026
🏷️ Format: Keynote Launch
