Why this is in the vault
A single-argument follow-on to Packkildurai's "self-organizing data stack" piece, arguing that in an agent-built data system the data model's job shifts from "diagram the schema" to "make the agent's interpretation of a request inspectable before it builds" — directly relevant to RDCO's dispatch-review and independent-verification patterns.
Mapping against Ray Data Co
The piece's worked example — an agent asked to find "customers who responded to a campaign," which silently picks a click-based definition instead of the purchase-based one the marketing team actually meant, builds a working pipeline, passes every test, and still answers the wrong question — is the data-layer version of exactly the failure verify-dispatch and "structure before research dispatch" exist to catch: a producer executes on its own unstated interpretation of an ambiguous ask before anyone confirms the interpretation was right. Packkildurai's fix is the same shape as RDCO's: surface the interpretation (what it inferred vs. what a domain owner confirmed, what it couldn't resolve) before implementation, not as a post-hoc summary that may not even match what got built. He makes the same point RDCO already encodes operationally — "teams will need ways to check the definitions and decisions against the code, data, and tests, not just read the agent's summary and assume they match" is the identical worry behind routing vault writes and dispatch prompts through a fresh-eyes critic that never trusts the producer's own account of its work.
Second connection: this is an explicit sequel to Packkildurai's own [[2026-10-14-data-engineering-weekly-self-organizing-data-stack]] (already filed), which argued orchestration engines — not engineers — should own the DAG, gated by verification against business invariants rather than syntax. This piece answers the natural follow-up question that framing raises: if an engine/agent owns more of the design, what does a human need to see to trust it? His answer (legible, inspectable interpretation decisions, tied to an immutable decision log) is the same "humans own meaning, agents own mechanics" split RDCO already treats as a hard architectural bet, now extended one layer up from pipeline execution to requirements interpretation.
Honest limit: no named product, benchmark, or deployed system here — it's conceptual argument, same as the piece it follows. Value to RDCO is as reinforcing evidence for an already-held position, not new tooling or a build blueprint.
The core argument
Packkildurai argues that as coding/data agents take over more pipeline design and implementation (his "Agentic Data Development Lifecycle," ADDLC), the answer to "do we still need data models" is yes — but the model's role changes. Historically a data model exists so a human team can agree on business meaning (what a row represents, how returns affect sales, which address version matters) before implementation. When an agent does the implementation, people still need that same shared meaning, but now also need visibility into how the agent interpreted an ambiguous request, which relationships it inferred versus had confirmed, and where it lacked enough information to decide — because an agent can produce a plausible explanation of its work that doesn't match what it actually built. He frames this as two connected views that must stay linked: the dimensional/technical model that organizes data for analysis, and a business-facing view of the agent's interpretation that a domain owner can inspect, correct, and confirm. He's agnostic on technique — Kimball dimensional modeling, semantic layers, or an ontology (citing Palantir's AIP) can each serve this depending on the workload — but insists the investment in any of them needs a clear owner and purpose, not adoption for its own sake.
Related
- [[2026-10-14-data-engineering-weekly-self-organizing-data-stack]]
- [[2026-01-09-trevin-chow-agent-orchestration-thesis]]
- [[2026-08-14-data-engineering-weekly-agent-ontology]]
- [[feedback_structure_before_research_dispatch]]
- [[feedback_verification_independent_worker_pattern]]