Typedef co-founder Yoni Michael spoke at an AI Engineer workshop, and his argument was straightforward. Coding agents don't fail on data because the models are weak. They fail because context is missing. He then demoed the Data Context Layer his team built to fix it. Michael's bet is that agents will become core operators of warehouses and pipelines, which means they need a reasoning layer that mirrors how the data team that built the system understands it end to end.
Valid SQL is not enough.
Michael said the missing context is scattered everywhere. The warehouse holds rows, transformation code holds business logic, the BI tool holds metrics and dimensions, and docs and Slack hold tribal knowledge. An agent that only reads code re-derives context every prompt, probabilistic and bounded by the window. An agent with only database access via a Snowflake or Databricks MCP connection can sample rows but cannot tell what one row is supposed to mean. He called that meaning grain, what each row represents, and showed how a grain change breaks everything downstream even when queries compile. If a customer by month table silently becomes customer by week, every sum after it is wrong.
The layer makes four facts computable and queryable: grain, whether a join is safe, what a metric means, and what relates to what. Michael said the principle is deterministic first, LLM only for ambiguity. Typedef parses all transformation SQL down to its abstract syntax tree to compute grain and lineage straight from implementation, so the graph cannot drift. It then verifies join safety by profiling tables in the warehouse to check if a join key is actually unique and what cardinality will fire. The LLM comes in last, naming assets, summarizing model utility, and resolving human readable ambiguity on top of the deterministic graph.
That graph is the Data Context Layer itself. It is computed once and refreshed atomically on every commit, so it stays representative of production. Michael said it is a graph not a pile of markdown because the access pattern matters. The agent pulls exactly the subgraph it needs in one traversal with Cypher, instead of prompt stuffing a 500-model dbt project every session. Edges carry provenance, with typed relations like is materialized by and has semantic definition, connecting physical tables, logical models, and semantic views and dashboards from sources to BI. The same idea applies outside SQL. As a daily.dev summary of the talk notes, two Rust types both called PostingList break different code. A text search finds both but flattens them, while a context graph preserves their distinct identities and blast radii, which is why graph traversal beats keyword retrieval for impact analysis.