Core Concepts

Scout provides three core primitives for building AI-powered automations. Understanding how they differ—and how they connect—is key to building effectively.

The Three Primitives

Workflows vs Agents

This is the most important distinction in Scout:

When to use workflows:

• Processing data through multiple steps
• Calling APIs and transforming responses
• Conditional logic and branching
• Scheduled or triggered automations

When to use agents:

• Conversational interfaces (chat)
• Tasks requiring reasoning and judgment
• Multi-step tool use where next step depends on previous result
• When you want the LLM to decide what to do

How They Connect

Workflows and agents can invoke each other:

• Slack triggers can execute workflows OR agents directly
• Copilot typically triggers a workflow that routes to an agent
• Workflows can call agents via the Message Agent block
• Agents can call workflows as tools

Workflow Execution Model

Workflows execute as a directed acyclic graph (DAG):

1. Blocks with no dependencies run first
2. Independent blocks run in parallel (not sequentially)
3. A block only runs when all its predecessors complete
4. Conditional blocks can skip downstream branches

In this example:

• LLM Block and HTTP Block run in parallel (both depend only on Input)
• Conditional waits for both to complete before running
• Only one branch (E or F) executes based on the condition

State & Data Flow

Workflow State

Each block stores its output in a shared state object. Access data using Jinja templates:

1
{# Access input fields by their ID #}
2
{{ inputs.transcript }}
3
4
{# Access block outputs using block_id.output #}
5
{{ llm.output }}
6
{{ http_block.output.data | tojson }}

Agent Sessions

Agents maintain conversation history per session:

• Each user-agent conversation has a unique session
• Messages accumulate across turns
• Context is preserved until session ends

Databases & RAG

Databases store documents with vector embeddings for semantic search:

1. Ingest: Upload documents (text, PDFs, web scrapes)
2. Embed: Automatic vector embedding generation
3. Query: Search by meaning, not just keywords
4. Use: Pass retrieved context to LLM blocks or agents

Next Steps