Agent Workflow

Every agent in Discover is powered by a workflow — a visual graph of connected components that defines how a user’s query is processed from input to answer. Workflows are configured through the no-code canvas editor and can be customized to fit your specific use case.

Here’s how a sample agent looks like:

Workflow Structure

A typical agent workflow follows this general pattern:

User Input → Security / Validation → Context Retrieval → Processing → Answer

Each step is represented by a component (node) on the canvas. Components are connected by edges that define the flow of data. See Agent Components for a full list of available components.

Agent Templates

Discover provides three built-in agent templates, each with a preconfigured workflow optimized for its use case:

Template	Best For
Database Agent	Natural language queries over relational databases (PostgreSQL, MySQL, MariaDB)
API Agent	Fetching and processing data from external REST APIs
Coordinator Agent	Entrypoint that orchestrates multiple sub-agents

Common Workflow Features

Security Checks

All agent workflows include security validation to detect and block malicious inputs — including SQL injection, prompt injection, tautologies, and schema enumeration attempts — before any processing occurs.

Date Handling

Agents automatically convert relative date expressions (e.g., “last 30 days”, “this quarter”) into absolute date ranges. Dates are never hardcoded — they are always calculated dynamically at query time.

LLM Configuration

Agent workflows use precision-focused model settings to ensure consistent, deterministic outputs:

• Low temperature for factual, reproducible responses
• Chain-of-thought reasoning before generating SQL or API payloads
• Retry and self-repair loops on failure