AI Agents

Chat-based AI assistants with tool access — schema browsing, query suggestions, and recipe guidance

Chat Agents (v0.11.0+)

AI agents are interactive chat assistants that operate within the Workbench. Unlike the AI-guided recipe builder (which follows a structured concierge flow), agents are open-ended — you ask questions, they answer and can take actions on your behalf. Agents have access to your datasets, database schemas, and recipes.

Agent Capabilities

Agents in the Workbench can: browse database schemas and list tables, preview data from any table, suggest SQL queries based on natural language questions, explain recipe stages and suggest improvements, answer statistical questions about your data, and execute read-only queries against your databases. Agents never write data unless you explicitly confirm.

AI Providers

Agents use the same AI providers configured for the recipe assistant. You can switch providers per session — cloud providers (DeepSeek, OpenAI, Anthropic) for fast reasoning, or a local LLM for complete data privacy. See AI Providers for configuration details.

Audit Trail

All agent interactions are logged: your questions, the agent's responses, and any actions the agent took. The audit trail is visible in the chat history and can be exported. This ensures that AI-assisted decisions are traceable and reviewable — an important control for regulated environments.

Agent Tools

Agents use a tool-calling architecture. When you ask "What's the average sales by region?", the agent calls the schema browser to find the relevant table, then the query runner to execute the aggregation, then formats the results. Each tool call is logged in the audit trail.

Agents are read-write with confirmation. Read-only operations (schema browsing, data preview, SQL queries) run without confirmation. Write operations (modifying recipes, deleting datasets, schema changes) require explicit user approval before the agent can proceed.
← AI Providers 📋 Contents Database Connections →