CONCEPT Cited by 1 source

Asset-first agent design¶

Asset-first agent design is the principle that an LLM agent should surface existing, trusted assets — curated tables, queries, dashboards, metric definitions — before generating new artifacts from scratch.

(Source: sources/2026-03-06-pinterest-unified-context-intent-embeddings-for-scalable-text-to-sql.)

The principle¶

Pinterest states it directly for the Analytics Agent:

"A core design principle is the asset-first approach: the agent should surface existing, trusted assets — tables, curated queries, dashboards, metric definitions — before generating new SQL."

Why it matters¶

Two reasons:

Reuse drives consistency. When two analysts ask about retention two weeks apart, they should get the same metric definition — not two independently-generated (and subtly different) SQL queries. Asset-first retrieval routes both to the same curated asset.
Existing assets encode validated expertise. Curated queries + dashboards + metric definitions already survived review, validation, and production use. They are encoded domain expertise; an LLM generating from scratch loses this accumulated signal.

Implementation at Pinterest (today)¶

Asset-first is currently implemented for tables and queries:

Table discovery — Analytics Agent surfaces Tier-1 tables first (governance-aware ranking).
Query discovery — Analytics Agent retrieves semantically similar past queries via unified context-intent embeddings and offers them for direct reuse before falling back to SQL generation.

Roadmap: progressive asset coverage¶

Pinterest's named roadmap explicitly widens the asset set:

Dashboards
Datasets
Metric definitions
Curated query libraries
Workflow artifacts

"Surfacing trusted, pre-existing answers before generating new queries, and making the full breadth of Pinterest's analytical assets discoverable through natural language."

Contrast: generate-first agent design¶

A generate-first agent treats every user question as a fresh generation task. This is easier to build but harder to govern, scales poorly across analysts (each gets a different answer), and discards the compounding value of a growing curated library.

Seen in¶

sources/2026-03-06-pinterest-unified-context-intent-embeddings-for-scalable-text-to-sql — canonical wiki definition; Pinterest Analytics Agent.