2-column Layout
A familiar two-column agent-building design was used to align with other products in the Google ecosystem.
Google Cloud Agent Builder
Summary
Custom data agents promised to help experts answer more questions across their organizations, but raw BigQuery data lacked the business context those agents needed to be accurate. I led the design of an authoring experience where experts could explain what their data meant, test the agent's answers, and publish it for colleagues who did not know SQL. In early private-preview evaluations, the structured context collected by the builder improved agent accuracy by roughly 10-20%.
Challenge
Make a complex agent-building experience easy for data analysts, so they could build agents with richer enterprise context, improve answer accuracy, and share them with colleagues.
Outcome
- +10-20%Higher agent accuracy from structured context during early evaluation.
My Contribution
- Reframed agent creation from step-by-step setup into an iterative build-and-test workflow.
- Created a reusable pattern for reviewing AI-generated content at scale that was adopted by the Google Cloud Platform design system.
- Defined lifecycle states that let users safely develop and publish agents.
2-column Layout
A familiar two-column agent-building design was used to align with other products in the Google ecosystem.
Metadata
Users can drill into an agent’s data sources to add rich metadata that boosts agent understanding of the data. AI-assisted context generation is provided to speed up the process.
SQL as Context
Users provide “golden” SQL queries to guide agents as they write SQL queries.
Suggested Queries
The system generates SQL queries based on the selected data sources. Users can verify and add the suggestions to their agent’s context.
Glossary Terms
Users can add terms specific to their business to increase agent understanding of internal vocabulary.
Testing
Once context is added, users can test their agent and iterate on their settings.
Agent Hub
Once saved or published, the agent appears in the agent hub, where users can start conversations or share their agent across platforms.
Data Studio
The agent can be shared onto less technical surfaces, like Google Data Studio, where their colleagues work.
Outcome
Currently in private preview, initial evaluations showed a promising +10-20% increase in agent accuracy when using the structured context of the builder.