Skip to main content

​
What is an agent?

An agent in Interact is an AI assistant configured specifically for your data and your questions. Unlike a generic AI chatbot, an agent knows:
  • Which data it can access — via your BigQuery connector
  • What your data means — through context you provide (table descriptions, business rules, metric definitions)
  • What kind of analysis to do — based on how you configure its subagents and tools
You can build multiple agents for different purposes — one for campaign performance, one for budget tracking, one for anomaly detection.

​
Agent architecture: orchestrator + subagents

Every agent follows the same structure:

​
Orchestrator

The orchestrator is the “manager” agent. It:
  • Reads and interprets your question
  • Decides which subagents to call and in what order
  • Assembles the final response from their outputs
You configure the orchestrator with a system prompt that describes the agent’s purpose and any business rules it should follow.

​
Subagents

Subagents are specialists that the orchestrator calls to do specific work. Currently, Interact supports two built-in subagent types:

Database Query Agent

What it does: Generates and runs SQL queries against your BigQuery connector.When to use it: Any time your agent needs to retrieve data from your warehouse.

Data Analysis Agent

What it does: Interprets query results — computing summaries, identifying trends, formatting charts.When to use it: When you want the agent to reason about the data, not just return raw rows.
Most agents need both subagents working together: the Database Query Agent fetches the data, and the Data Analysis Agent makes sense of it.

​
Agent context

Context is the information you give the agent about your organisation and data model. The more relevant context you provide, the more accurate the agent’s answers will be. There are two levels of context:
LevelScopeExample
Organisation contextShared across all agents in your workspace”Our main markets are NL, BE, and DE. Campaign IDs follow the format CMP-YYYY-XXX.”
Agent contextSpecific to one agent”This agent analyses paid search campaigns. Focus on the report_channels__performance dataset.”

​
Per-agent AI models

Each agent (orchestrator and subagents individually) can use a different AI model. This lets you balance cost and quality:
  • Orchestrator — Use a more capable model for complex reasoning and coordination
  • Database Query Agent — A faster, cheaper model is often sufficient for SQL generation
  • Data Analysis Agent — A capable model helps with nuanced interpretation and chart generation
If you don’t set a model override, the agent uses your organisation’s default model.

​
Prompt versioning

Every change you make to an agent’s prompt is saved as a new version. You can view the history and restore any previous version at any time. Learn more about prompt management →

​
Agent visibility and access

You control who can see and interact with each agent:
  • Private — Only visible to you
  • Organisation — All members of your workspace can use it
  • Scheduled only — Not accessible via chat; runs only on schedule
Learn how to build an agent →