Discovery: AI Sandbox with Full Database Access

[2witstudios]

Overview

Design the AI sandbox environment that gives agents full SQL and code execution access within tenant boundaries.

Key Questions

1. How does AI connect to tenant's PostgreSQL? (Connection pooling, direct, etc.)
2. What code execution environment? (Node.js, Python, both?)
3. Shell access scope and limits?
4. Error handling and iteration (AI sees error, fixes query)?
5. Token limits for query results?

Technical Options

• E2B code execution
• Modal serverless
• Custom Docker sandbox
• Firecracker microVM

Deliverable

Sandbox architecture with security model and execution flow.

Current State

The integration/tool execution system is in production (packages/lib/src/integrations/), supporting HTTP, GraphQL, function, and chained execution. This discovery is about extending that to full database and code execution access within isolated tenant infrastructure.

Related

• Depends on: Multi-tenant Infrastructure Isolation (Discovery: Multi-tenant Infrastructure Isolation Architecture #772)
• Enables: AI agents with full programmatic access to tenant data