When skills meet scale: an agent assembly line for dbt

At a dbt Summit 2026 breakout session titled "When skills meet scale: an agent assembly line for dbt," Jess Zhang, senior data engineer at Riot Games, walked through how her team used AI agents to rebuild a production data warehouse to engineering standards, at scale, during a real Wild Rift event.

In the old software development lifecycle, business logic lived inside massive PySpark jobs where orchestration, configuration, and transformation code were all braided together. Zhang described three costs of that setup: an "unboxing tax" to understand any one metric, a "testing tax" from inconsistent validation, and an "exclusivity tax" that kept domain experts from contributing without learning the entire tooling stack.

Zhang's team piloted a new agentic approach on Echoes on the Rift, a nostalgia event launched August 7, 2026 that linked players' mobile Wild Rift and PC League of Legends accounts across five years of history and more than 60 cross-title metrics.

The single 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 just to calculate basic battle stats like total kills and damage dealt.

The system Zhang built around this problem uses a layered context architecture: a system prompt built by engineers defines the agent's persona and tools, an injected "playbook" layer built jointly by engineers and analysts encodes naming conventions and schema contracts, and a task prompt lets an analyst describe what they need in a single sentence.

That context feeds what Zhang called an assembly line: five stations that 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 the feedback loop fast, the system checks generated models at three levels: syntax, whether the SQL is well-formed, compilation, whether referenced tables and columns actually exist, and execution, where the query runs against real data to check assumptions like null rates and row counts.

Zhang credited dbt v2 with running those compilation-level type checks against the catalog in seconds rather than requiring a full warehouse execution for every small error.

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

Zhang argued the real shift is in the cost curve: in the old model, the 200th model cost about as much engineering time as the first, while 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.

She was careful to note the approach does not remove engineers from the loop. 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.

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

Watch the full session, plus every keynote and breakout from dbt Summit 2026, on demand. Watch dbt Summit 2026 on demand →