Riot Games Uses an AI Agent Assembly Line to Rebuild a dbt Warehouse

From braided PySpark jobs to an agentic workflow

In a breakout session titled "When skills meet scale: an agent assembly line for dbt," Zhang described how her team rebuilt a production data warehouse to engineering standards using AI agents. The starting point was the old development lifecycle, in which business logic lived inside massive PySpark jobs where orchestration, configuration, and transformation code were all braided together.

Zhang named three costs of that setup. The first was an "unboxing tax" to understand any single metric. The second was a "testing tax" caused by inconsistent validation. The third was an "exclusivity tax" that kept domain experts from contributing unless they learned the entire tooling stack.

The team piloted the new approach on Echoes on the Rift, a nostalgia event launched August 7, 2026. It linked players' mobile Wild Rift and PC League of Legends accounts across five years of history and more than 60 cross-title metrics. The League of Legends PC game event table alone held about 18 terabytes of data, roughly 129 billion raw game events that had to be combed through to calculate basic battle stats such as total kills and damage dealt.

Layered context and five stations

The system rests on a layered context architecture with three parts:

• A system prompt built by engineers that defines the agent's persona and tools.
• An injected "playbook" layer, built jointly by engineers and analysts, that encodes naming conventions and schema contracts.
• A task prompt that lets an analyst describe what they need in a single sentence.

That context feeds what Zhang called an assembly line of five stations. They discover source columns, profile their value ranges, author a silver model with standard naming and join patterns, build a gold-layer model, and validate the result before a human reviews the pull request.

To keep feedback fast, generated models are checked at three levels. Syntax checks whether the SQL is well-formed. Compilation checks whether referenced tables and columns actually exist. Execution runs the query against real data to test assumptions such as null rates and row counts. Zhang credited dbt v2 with running compilation-level type checks against the catalog in seconds, instead of requiring a full warehouse execution for every small error.

Benchmark and the cost curve

Zhang compared Echoes on the Rift with a similar event she built by hand the previous year. The manual project took two months of fully protected, dedicated focus time. Echoes on the Rift took about two and a half months, even though she stayed on the on-call rotation, absorbed shifting scope, and built the underlying agent system from scratch in parallel.

Her main argument concerned the cost curve. In the old model, the 200th model cost about as much engineering time as the first. In the agentic model, the upfront cost of building the playbook and evaluation system is paid once, and the marginal cost of each additional model drops toward token and compute cost.

Engineers stay in the loop

Zhang was careful to say the approach does not remove engineers. It shifts their effort from writing boilerplate SQL toward data modeling precision, review, and separating the mechanics of SQL from the semantics of the business logic.

One open problem remains. During Q&A, Zhang acknowledged that preventing multiple contributors from building redundant models is unsolved. She manages it today through PR review and instructions that tell the agent to check for existing models before building new ones.

For data teams, the session points to a model where analysts can describe needs in plain language while engineers concentrate on playbooks, evaluation and review. The full session is available on demand through dbt Summit 2026.