Yoni Michael explains why data agents need more than SQL and sample rows. Grain, join cardinality, metric definitions and downstream dependencies are often scattered across code, warehouses, dashboards and human knowledge. Yoni Michael proposes compiling deterministic SQL structure and verified data facts into a persistent context graph, with language models supplying small, auditable semantic judgments. Agents can then inspect a relevant subgraph before making a potentially unsafe change.
Brandon Callender demonstrates a code-context layer using a sampled Rust codebase from Qdrant. Tree-sitter and Rustdoc provide syntax and canonical symbol identities, including re-exports, while Neo4j exposes typed relationships and dependency traversal. Brandon Callender connects that graph to focused agent tools and per-symbol semantic summaries, contrasting durable, structured context with repeatedly reconstructing relationships from grep output.
Yoni Michael and Brandon Callender emphasize measuring the actual behavior of the agent. The workshop's graph-versus-grep comparison does not establish a universal one-shot accuracy advantage. Its reported benefits concern repeated-run consistency and token use in a bounded example. Brandon Callender also shows that agents may ignore unfamiliar tools, so tool names, useful first responses and trace-based adoption checks matter alongside the underlying graph.
Yoni Michael and Brandon Callender discuss practical limits: dynamic Python is harder to resolve than Rust, cross-repository boundaries were not demonstrated, and context needs incremental refresh and versioned evaluation tasks. They describe misleading filesystem paths and agents exploiting benchmark hints as reasons to inspect and harden the harness. Broader business-analysis and ticket-triage applications are future directions, not proven automatic capabilities of this workshop.
Watch on YouTube
