Agent Blocks in Workflows

Bring AI intelligence to your automated workflows by using agents as workflow blocks.

What are Agent Blocks?

Agent Blocks let you run AI agents as steps within a workflow. This combines the reliability of predefined workflows with the adaptability of AI agents, giving you the best of both worlds.

Instead of rigid if/then logic, an Agent Block can:

• Analyze data and make intelligent decisions
• Handle variations and unexpected inputs
• Extract information from unstructured text
• Generate personalized content
• Route work to the right destination

When to Use Agent Blocks

Processing Unstructured Data

When a webhook receives data that doesn’t follow a predictable format:

Trigger: New email arrives
   ↓
Agent Block: Analyze email, extract sender intent, urgency and key information
   ↓
Condition: Route based on agent's analysis
   ├─ Urgent → Create high-priority ticket
   ├─ Sales inquiry → Forward to CRM
   └─ General → Auto-respond and archive

Intelligent Routing

Let an agent decide where data should go:

• Categorize support tickets by topic and assign to the right team
• Evaluate leads and route to appropriate sales rep
• Analyze feedback and tag for product, engineering or marketing

Content Generation

Create personalized content as part of an automation:

• Generate custom email responses
• Create meeting summaries from transcripts
• Draft reports based on collected data

Research and Enrichment

Have an agent research and add context:

• Webhook receives a new contact → Agent researches company and enriches record
• Lead comes in → Agent finds relevant news and adds talking points
• Customer mentioned → Agent pulls recent activity for context

How Agent Blocks Work

Input Configuration

Agent Blocks receive data from:

• Workflow trigger — The initial event (webhook payload, schedule, etc.)
• Previous blocks — Output from earlier workflow steps
• Static values — Fixed parameters you define

Agent Selection

Choose any agent you’ve built in Scout:

• Select from your agent library
• Pass context and instructions specific to this workflow
• Configure response format (text, JSON, structured data)

Output Handling

The agent’s response becomes available for subsequent blocks:

// Access agent output in later blocks
{
  "agent_result": {
    "category": "technical-support",
    "urgency": "high",
    "suggested_response": "Based on the error...",
    "assign_to": "engineering-team"
  }
}

Example: Meeting Follow-up Workflow

This workflow demonstrates how Agent Blocks create intelligent automation:

Trigger: Webhook (meeting transcript received from video conferencing)
   ↓
Agent Block: Analyze Transcript
   Input: meeting_transcript
   Instructions: "Extract action items, decisions made and follow-ups. 
                  Identify owners and deadlines."
   ↓
Action Block: Create Tasks
   For each action item → Create task in project management tool
   ↓
Agent Block: Draft Follow-up Email
   Input: meeting_summary, attendees, action_items
   Instructions: "Write a professional follow-up email summarizing 
                  the meeting and highlighting action items."
   ↓
Action Block: Send Email
   Send draft via Gmail to all attendees

The workflow is reliable and repeatable, but the agents handle the nuance of understanding each unique meeting.

Example: Lead Intelligence Workflow

Trigger: Webhook (new lead from website form)
   ↓
Agent Block: Research Company
   Input: lead_company_name
   Instructions: "Research this company. Find their industry, 
                  size, recent news and potential pain points 
                  our product could address."
   ↓
Action Block: Create CRM Record
   Create contact with enriched data
   ↓
Agent Block: Generate Outreach
   Input: research_findings, lead_info
   Instructions: "Write a personalized outreach email addressing 
                  their likely needs based on the research."
   ↓
Condition: Check Lead Score
   If high-value → Notify sales rep immediately
   Otherwise → Add to nurture sequence

Best Practices

Keep Agents Focused

Each Agent Block should do one thing well:

• ✅ “Extract the purchase intent from this message”
• ❌ “Analyze everything about this customer and decide what to do”

Chain multiple focused Agent Blocks instead of one complex agent.

Define Clear Outputs

Structure your agent instructions to return predictable formats:

Instructions: "Analyze this support ticket. Return JSON with:
- category: one of [technical, billing, feature-request, other]
- urgency: one of [low, medium, high, critical]
- summary: 1-2 sentence summary
- suggested_action: what should happen next"

This makes the output easy to use in subsequent workflow blocks.

Handle Edge Cases

Agents are powerful but not infallible. Add fallback logic:

Agent Block → Condition Block (was output valid?)
   ├─ Yes → Continue workflow
   └─ No → Route to human review queue

Use Agent Instructions as Configuration

Treat agent instructions like workflow configuration. Be specific about:

• What input the agent receives
• What format the output should be
• Any constraints or preferences

Limitations

• Agent Blocks have the same timeout limits as chat interactions
• Large or complex tasks should be broken into smaller Agent Blocks
• For purely deterministic logic, use Condition Blocks instead

Next Steps

• Workflows Overview — Learn more about workflow fundamentals
• Agents — Build agents to use in your workflows
• Getting Started — Start building your first workflow

Built with ❤️ by Scout OS