lindy-core-workflow-a

# Lindy Core Workflow A: Agent Creation

## Overview
Complete workflow for creating, configuring, and testing Lindy AI agents.
Agents consist of four components: **Prompt** (behavioral instructions),
**Model** (AI engine), **Skills** (available actions), and **Exit Conditions**
(completion criteria).

## Prerequisites
- Lindy account at https://app.lindy.ai
- Use case defined (support bot, email triage, data processor, etc.)
- Required integrations identified (Slack, Gmail, Sheets, etc.)

## Instructions

### Step 1: Define Agent Specification
Before building, document:
- **Purpose**: What the agent does (one sentence)
- **Trigger**: What wakes it up (webhook, email, schedule, chat, etc.)
- **Actions**: What it does (send email, update sheet, post to Slack, etc.)
- **Model**: GPT-4 (smart, expensive), Claude Sonnet (balanced), Gemini Flash (fast, cheap)

### Step 2: Create the Agent
**Option A — Natural Language (recommended)**:
1. Click **New Agent** at https://app.lindy.ai
2. Describe your agent in plain English:
   ```
   When a customer emails support@company.com, classify the email as
   billing/technical/general, draft a response using our knowledge base,
   and post the classification to #support-triage in Slack
   ```
3. Agent Builder auto-generates trigger + action nodes

**Option B — Manual Build**:
1. Click **New Agent** > **Start from scratch**
2. Add trigger: Click **"+"** > Select trigger type
3. Add actions: Click **"+"** > Search for action
4. Connect nodes by dragging arrows between steps

### Step 3: Configure the Prompt
Open **Settings > Prompt**. Structure it with clear sections:
```
## Identity
You are a customer support classifier and responder for [Company].

## Instructions
1. Read the incoming email carefully
2. Classify into: billing, technical, or general
3. Search the knowledge base for relevant answers
4. Draft a professional response using the KB results
5. If no KB match found, escalate to human

## Constraints
- Never promise refunds or credits without human approval
- Keep responses under 200 words
- Always include ticket reference number
```

**Prompt best practices** (from Lindy docs):
- Use action-oriented language (imperatives)
- Break complex logic into numbered steps
- Include few-shot examples for consistent formatting
- Add constraints to prevent unwanted behavior

### Step 4: Add Conditions (Branching Logic)
1. Click **"+"** > **Condition**
2. Write natural language condition: `"Go down this path if the email is about billing"`
3. Add multiple branches for different classifications
4. Enable **"Force the agent to select a branch"** for deterministic routing

### Step 5: Configure Actions
For each action, set field modes:

**Auto mode** — Agent infers the value from all previous step data:
```
Best for: predictable mappings where field names align
```

**AI Prompt mode** — Give natural language instructions:
```
Summarize the email in 2 sentences, then include the classification.
Reference: {{email_received.body}}
```

**Set Manually mode** — Exact value, no AI:
```
Channel: #support-triage
```

### Step 6: Add Knowledge Base (Optional)
1. Go to **Settings > Knowledge Base**
2. Add sources: PDF, DOCX, Google Drive, Notion, websites
3. Configure search:
   - **Max Results**: 4 (default) to 10
   - **Search Fuzziness**: 0 (keyword) to 100 (semantic, recommended)
4. Agent auto-searches KB when relevant to the task

### Step 7: Test the Agent
1. Use the **Test** button in the workflow editor
2. Provide sample input matching your trigger type
3. Review each step's output in the task detail view
4. Iterate on prompt and action configuration

## Trigger Types Reference

| Trigger | Use Case | Configuration |
|---------|----------|---------------|
| **Webhook Received** | External API calls | URL + secret key |
| **Email Received** | Inbox automation | Gmail/Outlook + label filters |
| **Schedule** | Recurring tasks | Cron-style: daily, weekly, custom |
| **Chat Message** | Interactive bot | Lindy Chat or Embed widget |
| **Slack Message** | Team automation | Channel + keyword filters |
| **Agent Message** | Multi-agent delegation | Receives from other Lindies |
| **Calendar Event** | Meeting automation | Minutes offset (-30 = 30 min before) |
| **Form Submission** | Lead capture | Connected form integration |

## Action Categories

| Category | Actions |
|----------|---------|
| **Email** | Send Email, Draft Reply, Search Inbox, Add Label |
| **Slack** | Send Channel Message, Send DM, Thread Reply |
| **Sheets** | Update Spreadsheet, Get Document |
| **Calendar** | Create Event, Reschedule, Cancel |
| **Knowledge** | Search Knowledge Base, Resync KB |
| **Code** | Run Code (Python/JS in E2B sandbox) |
| **Web** | HTTP Request, Web Search, Website Crawler |
| **Memory** | Read/Create/Update/Delete Memory |
| **Phone** | Make Call, Transfer Call, End Call |
| **Agent** | Agent Send Message (delegation) |

## Error Handling

| Error | Cause | Solution |
|-------|-------|----------|
| "No trigger configured" | Agent has no trigger | Add at least one trigger node |
| Action fails silently | Wrong field mode | Switch to AI Prompt or Set Manually |
| KB returns no results | Fuzziness too low | Increase to 100 (semantic search) |
| Condition always picks same path | Ambiguous prompt | Make conditions more specific |
| Agent loops indefinitely | No exit condition | Add measurable exit criteria |

## Resources
- [Lindy Introduction](https://docs.lindy.ai/fundamentals/lindy-101/introduction)
- [Triggers Documentation](https://docs.lindy.ai/fundamentals/lindy-101/triggers)
- [Actions Documentation](https://docs.lindy.ai/fundamentals/lindy-101/actions)
- [Prompt Guide](https://docs.lindy.ai/fundamentals/lindy-101/prompt-guide)

## Next Steps
Proceed to `lindy-core-workflow-b` for triggers, automation, and multi-agent delegation.