Yoni Michael and Brandon Callender from Typedef present the Data Context Layer, a graph-based infrastructure that integrates diverse data sources and code semantics to provide AI agents with a persistent, computable understanding of data and code, enabling reliable reasoning, impact analysis, and safe transformations. They demonstrate practical implementation using code parsing tools and discuss strategies to encourage agent tool usage, emphasizing modularity, scalability, and the complementary role of semantic understanding to enhance AI-driven data management and decision-making.
The presentation by Yoni Michael and Brandon Callender from Typedef centers on the concept of the Data Context Layer, a foundational infrastructure designed to empower AI agents to effectively operate within complex data environments. Yoni introduces the challenge that while AI models continue to advance, the real bottleneck lies in providing agents with the necessary context to understand and safely manipulate data systems. He emphasizes that agents often fail not due to lack of coding ability but because they lack a persistent, computable understanding of data semantics such as grain, join safety, and metric definitions. Typedef’s solution is to build a deterministic, queryable graph-based context layer that integrates information from diverse sources like data warehouses, transformation code, BI tools, and documentation, enabling agents to reason about data changes, impact analysis, and safe transformations reliably.
Brandon’s portion of the workshop translates these concepts into a hands-on coding example using the Qdrant codebase. He demonstrates how to build a context layer for a coding agent by extracting structured information from code using tools like Tree-sitter and Rust’s compiler outputs. This process involves multiple levels of context: from raw text search (level one), to semantic search (level two), and ultimately to a traversable graph (level three) that captures the relationships and identities of code symbols. By constructing this graph and integrating it with an agent, they enable the agent to perform complex reasoning tasks such as impact analysis on code changes, which would otherwise require extensive manual inspection.
A key insight from the workshop is the challenge of getting AI agents to actually use the sophisticated tools and context provided to them. Despite having access to a rich graph-based context, agents often default to simpler, familiar methods like text search due to their training and inherent reflexes. Brandon discusses strategies to encourage tool usage, such as designing tools that mimic familiar interfaces (e.g., naming a graph query tool “search” and formatting responses like grep) and ensuring tools provide immediate, useful results to maintain agent trust. He also highlights the importance of rigorous evaluation frameworks (eval-driven development) to measure not only correctness but also consistency and tool adoption rates, which are critical for building reliable agent workflows.
The discussion also touches on practical considerations such as incremental updates to the context graph to handle large and evolving codebases or data architectures efficiently. Both speakers acknowledge the complexity of scaling these systems and the necessity of modular, language-agnostic approaches to parsing and indexing. They note that while Rust provides helpful compiler outputs for deterministic analysis, other languages like Python pose greater challenges requiring additional tooling and heuristics. The integration of semantic understanding through LLMs is positioned as a complementary layer that summarizes and classifies code or data assets, enhancing the agent’s ability to interpret intent and meaning beyond structural relationships.
Finally, the presenters situate their work within the broader ecosystem of data management and AI tooling, clarifying that the Data Context Layer complements rather than replaces traditional data catalogs. Their vision is to unify disparate data sources, transformation logic, and business knowledge into a single, queryable graph that supports both technical teams and business users. This unified context enables faster, more accurate troubleshooting, impact analysis, and decision-making, ultimately bridging the gap between data engineers, analysts, and business stakeholders. The workshop concludes with an invitation to explore the open-source tools and frameworks Typedef has developed, emphasizing the ongoing challenge and opportunity in building trustworthy, context-aware AI agents for data systems.