Presentation: AI Agents to Make Sense of Data at OpenAI

2026-06-19 · Source: InfoQ · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, extended

Summary

OpenAI's Bonnie Xu presented Kepler, an internal AI data analyst agent designed to query over 600 petabytes of data across 70,000 datasets. Kepler addresses challenges like table discovery and complex SQL generation by leveraging a Multi-tool Co-optimization Process (MCP), automated code crawling for table metadata, and Retrieval Augmented Generation (RAG) for company context. The system also incorporates scoped semantic memory for continuous self-learning and utilizes AST-based LLM grading in its evaluation pipeline to ensure accuracy and prevent regressions. Kepler is available 24/7 via Slack, UI, or IDE, enabling users to ask complex data questions and receive detailed analyses, including charts and anomaly debugging.

Key takeaway

For MLOps Engineers building internal data tools, OpenAI's Kepler demonstrates a robust architecture for AI agents handling massive datasets. You should prioritize integrating comprehensive context beyond basic metadata, implement scoped memory for continuous learning, and establish AST-based LLM grading for evaluation to ensure accuracy and prevent regressions in complex query generation. This approach significantly boosts data productivity and user trust.

Key insights

OpenAI's Kepler agent uses advanced AI techniques to automate complex data analysis across vast datasets, improving data accessibility and accuracy.

Principles

• Context beyond table metadata is crucial.
• Memory enables continuous agent learning.
• Robust evals prevent model regression.

Method

Kepler employs MCP for interactive data exploration, automated code crawling for fresh table metadata, RAG for company context, and scoped semantic memory for corrections, all evaluated via AST-based LLM grading.

In practice

• Query 600+ petabytes via natural language.
• Debug data anomalies automatically.
• Generate complex SQL and charts.

Topics

• AI Agents
• Data Analysis
• Large Language Models
• Retrieval-Augmented Generation
• Data Platforms
• SQL Generation
• Evaluation Metrics

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• How OpenAI Built Its Data Agent — The reliability of AI agents at scale stems from strong data foundations and carefully engineered context, not complex agent architectures.
• AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery — AI agents struggle significantly with complex scientific literature discovery, achieving less than 10% accuracy on dedicated benchmarks.
• Stop Shipping on Vibes — How to Build Real Evals for Coding Agents — Robust AI evaluations are crucial for quantifiable shipping decisions, performance monitoring, and iterative improvement.
• Orion at Gravity: Trustworthy AI Analysts for the Enterprise — Agentic analytics combines semantic layers and broad context engineering to deliver trustworthy, actionable business insights.
• OpenAI unveils Workspace Agents, a successor to custom GPTs for enterprises that can plug directly into Slack, Salesforce and more — OpenAI's Workspace Agents enable persistent, code-executing AI workers across enterprise apps, evolving beyond session-based GPTs.
• OpenAI's AI data agent, built by two engineers, now serves 4,000 employees — and the company says anyone can replicate it — OpenAI's internal AI data agent demonstrates the transformative power of accessible data intelligence for enterprise operations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by InfoQ.