lindy-cost-tuning

# Lindy Cost Tuning

## Overview
Lindy uses a credit-based pricing model. Every task costs credits based on model
size, step count, premium actions, and duration. Cost tuning targets: model
right-sizing, agent consolidation, trigger optimization, and credit monitoring.

## Prerequisites
- Lindy workspace with billing access
- Multiple active agents to evaluate
- Dashboard access to review per-agent task history

## Credit Cost Reference

| Factor | Credits |
|--------|---------|
| Basic model task (Gemini Flash) | 1-2 |
| Mid-tier model (GPT-4o-mini, Claude Haiku) | 2-5 |
| Large model task (GPT-4, Claude Sonnet) | 5-10 |
| Premium model (Claude Opus) | ~10+ |
| Phone call (US/Canada) | ~20/minute |
| Phone call (international) | 21-53/minute |
| Premium actions (webhooks) | Additional per action |
| Minimum per task | 1 credit |

## Plan Costs

| Plan | Monthly | Credits | Per Extra Seat |
|------|---------|---------|----------------|
| Free | $0 | 400 | N/A |
| Pro | $49.99 | 5,000 | $19.99 |
| Business | $299.99 | 30,000 | Included |
| Enterprise | Custom | Custom | Custom |

## Instructions

### Step 1: Audit Agent Credit Consumption
For each active agent, collect:
1. **Task count** (last 30 days) — from Tasks tab
2. **Average credits per task** — total credits / task count
3. **Model used** — from agent settings
4. **Trigger frequency** — how often the agent fires

Create a cost audit table:
| Agent | Tasks/Month | Credits/Task | Model | Monthly Credits | % of Total |
|-------|------------|-------------|-------|----------------|-----------|
| Support Bot | 500 | 5 | Claude Sonnet | 2,500 | 50% |
| Lead Router | 200 | 2 | GPT-4o-mini | 400 | 8% |
| Report Gen | 30 | 10 | GPT-4 | 300 | 6% |

### Step 2: Right-Size Models
The highest-impact optimization. For each agent, ask:
> "Does this task actually need GPT-4/Claude, or would Gemini Flash work?"

| Current Setup | Optimized | Savings |
|--------------|-----------|---------|
| Email classify with Claude Sonnet (5 cr) | Gemini Flash (1 cr) | 80% |
| Data extract with GPT-4 (10 cr) | GPT-4o-mini (3 cr) | 70% |
| Simple routing with Claude Opus (10 cr) | Gemini Flash (1 cr) | 90% |

**Test the downgrade**: Run 10 tasks with the smaller model. Compare output quality.
Most classification, routing, and extraction tasks work identically on smaller models.

### Step 3: Consolidate Redundant Agents
Multiple single-purpose agents cost more than one multi-purpose agent:

Before (5 agents, 5 minimum credits per run):
```
Agent 1: Classify billing emails
Agent 2: Classify technical emails
Agent 3: Classify general emails
Agent 4: Draft billing responses
Agent 5: Draft technical responses
```

After (1 agent, 1 minimum credit per run):
```
Support Agent: Classify email → Condition (billing/technical/general)
  → Draft appropriate response → Send
```

**Cost impact**: Reducing from 5 agents to 1 saves minimum-credit overhead and
simplifies management.

### Step 4: Optimize Trigger Frequency
Credits are consumed every time a trigger fires. Reduce unnecessary triggers:

**Email Received**:
```
Before: Trigger on ALL emails (300/day) = 300 tasks
After:  Filter: label "support" AND NOT from "noreply@" (40/day) = 40 tasks
Savings: 87% fewer tasks
```

**Schedule trigger**:
```
Before: Every 15 minutes (96/day)
After:  Every 2 hours (12/day)
Question: Does this agent really need to run every 15 minutes?
```

**Slack trigger**:
```
Before: Any message in #general (200/day)
After:  Messages containing "@support-bot" (10/day)
Savings: 95% fewer tasks
```

### Step 5: Reduce Steps Per Task
Each action in a workflow costs credits. Eliminate unnecessary steps:
- Combine multiple LLM calls into one (see `lindy-performance-tuning`)
- Use Set Manually instead of AI Prompt for known values
- Remove debug/logging steps in production
- Simplify condition branches

### Step 6: Optimize Knowledge Base Usage
KB search costs credits per query. Optimize:
- Reduce Max Results from 10 to 4 (sufficient for most queries)
- Use specific query instructions to get relevant results in one search
- For small datasets (<100 entries), consider putting data directly in the prompt

### Step 7: Budget Monitoring Setup
1. Check credit usage weekly in **Settings > Billing**
2. Set internal alerts for high-consumption agents:
   - 50% of budget: Warning — review usage
   - 80% of budget: Alert — optimize or upgrade
   - 95% of budget: Critical — pause non-essential agents

### Step 8: Deactivate Idle Agents
Review agents monthly:
- No tasks in 30 days → Pause the agent
- No tasks in 90 days → Delete or archive
- Lindy only charges for active agent execution, not idle agents

## Monthly Cost Optimization Checklist
- [ ] Review per-agent credit consumption
- [ ] Identify agents using large models for simple tasks
- [ ] Check for redundant agents that could be consolidated
- [ ] Review trigger filter effectiveness
- [ ] Remove unused integrations from agents
- [ ] Verify no loops or runaway agent steps
- [ ] Compare actual spend to budget

## Error Handling

| Issue | Cause | Solution |
|-------|-------|----------|
| Unexpected credit spike | Trigger filter removed or loosened | Review and restore trigger filters |
| Agent consuming 10x normal | Looping agent step | Add exit conditions, check task history |
| Credits exhausted mid-month | Under-budgeted or spike | Upgrade plan or pause non-critical agents |
| Model downgrade hurts quality | Task needs larger model | Selectively upgrade only that step |

## Resources
- [Lindy Pricing](https://www.lindy.ai/pricing)
- [Lindy Documentation](https://docs.lindy.ai)

## Next Steps
Proceed to `lindy-reference-architecture` for production architecture patterns.