When skills meet scale: an agent assembly line for dbt
Riot Games built an AI agent assembly line that modeled 129 billion game events for a Wild Rift event without a human hand-coding every model.
Overview
This session explore moving beyond basic AI text generation toward practical agentic workflows, using dbt agent skills and the dbt v2 engine for secure multi-agent orchestration. See a workflow where read-only agents perform automated metadata discovery, convert legacy logic into dbt models, and generate PRs for human review, with guardrails keeping humans in control.
What you'll take away
- 01
An agentic "assembly line," five stations that discover columns, profile data, author a silver model, build a gold model, and validate the result, rebuilt a production data warehouse during a live Wild Rift event spanning 129 billion raw game events.
- 02
A layered context architecture, a system prompt for persona, a shared playbook for conventions, and a one-sentence task prompt from an analyst, is what let the agent work reliably at scale.
- 03
The real shift is in the cost curve: in the old model the 200th model cost as much engineering time as the first; in the agentic model the upfront cost is paid once and each additional model's marginal cost drops toward compute.
Session speaker
Jessica Zhang
Riot Games