Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
Data Science
Simon Willison's Blog·3 min read
datasette-agent 0.1a3
TL;DR
datasette-agent, a tool for querying databases using AI agents, lacked clear visibility into SQL operations and struggled with truncated responses. Version 0.1a3 added "View SQL query" buttons for both table results and collapsed tool calls, removed empty reasoning output, and improved handling of truncated data display.
✦ Why It Matters
Engineers can now debug AI-generated SQL queries directly and trust that data displays completely even with large result sets.
Key Takeaways
- Add SQL query inspection buttons to improve transparency of agent-generated queries.
- Remove empty reasoning output to reduce UI noise and improve user experience.
- Implement truncation-aware rendering to display complete tables despite length limits.
- Test agent behavior with large result sets to ensure data accessibility.
Full Summary
datasette-agent is a tool that lets users query databases conversationally through an AI agent that generates and executes SQL. Previous versions had usability gaps: users couldn't easily inspect the actual SQL queries being run, and when responses were truncated (cut off due to length limits), the table display would fail entirely.
Version 0.1a3 addresses these issues through three improvements. First, it adds clickable "View SQL query" buttons for both visible table results and collapsed SQL result tool calls (intermediate steps the agent takes).
Second, it stops displaying empty reasoning chunks, reducing UI clutter. Third, it implements smarter truncation handling so tables render completely for users even when the underlying SQL results exceed system limits.
These changes make the agent's decision-making process more transparent and ensure data remains accessible regardless of response length constraints.