Lindy Performance Tuning

Lindy Performance Tuning

Overview

Lindy agents execute as multi-step workflows where each step (LLM call, action execution, API call, condition evaluation) adds latency and credit cost. Optimization targets: fewer steps, smaller models, faster actions, tighter prompts.

Prerequisites

• Lindy workspace with active agents
• Access to agent Tasks tab (view step-by-step execution history)
• Understanding of agent workflow structure

Instructions

Step 1: Profile Agent Execution

In the Tasks tab, open a completed task and review:

• Total task duration: Baseline for improvement
• Per-step timing: Identify the slowest steps
• Credit consumption: Which steps cost the most
• Step count: Total actions executed per task

Common bottlenecks:

Step 2: Right-Size Model Selection

The single biggest performance lever. Match model to task complexity:

Rule of thumb: Start with the smallest model. Only upgrade if output quality is insufficient. Most classification and routing tasks work fine with Gemini Flash.

Step 3: Consolidate LLM Steps

Before (3 LLM calls, ~9 credits):

Step 1: Classify email (LLM)
Step 2: Extract key entities (LLM)
Step 3: Generate response (LLM)

After (1 LLM call, ~3 credits):

Step 1: Classify, extract entities, and generate response (single LLM prompt)

Consolidated prompt:

Analyze this email and return JSON with:
1. "classification": one of [billing, technical, general]
2. "entities": {customer_name, product, issue_type}
3. "draft_response": professional reply under 150 words

Email: {{email_received.body}}

Step 4: Use Deterministic Actions Where Possible

Replace AI-powered fields with Set Manually mode when values are predictable:

Each Set Manually field saves one LLM inference (~1 credit).

Step 5: Optimize Knowledge Base Queries

• Max Results: Set to the minimum needed (default 4, max 10)

• Search Fuzziness: Keep at 100 (semantic) unless precision matching needed

• Query mode: Use AI Prompt with specific instructions:

  Search for the customer's specific product issue.
  Focus on: {{extracted_entities.product}} {{extracted_entities.issue_type}}
  

  Not: "Search for relevant information" (too vague, wastes results)

Step 6: Optimize Trigger Filters

Prevent wasted runs with precise trigger filters:

Before: Email Received (all emails) → 200 runs/day → 600 credits
After:  Email Received (label: "support" AND NOT from: "noreply@")
        → 30 runs/day → 90 credits (85% savings)

Step 7: Use Agent Steps Judiciously

Agent Steps (autonomous mode) are powerful but expensive — the agent may take unpredictable paths and use more actions than a deterministic workflow.

Use Agent Steps when: Next steps are genuinely uncertain (complex research, multi-source investigation, adaptive problem-solving)

Use deterministic actions when: Steps are predictable (classify -> route -> respond)

When using Agent Steps:

• Limit available skills to 2-4
• Set clear, measurable exit conditions
• Include a fallback exit condition to prevent infinite loops
• Monitor credit consumption of first 10 runs to establish baseline

Step 8: Loop Optimization

For batch processing, configure loops for efficiency:

• Max Concurrent: Increase for independent items (parallel execution)
• Max Cycles: Always set a cap to prevent runaway processing
• Only pass essential data as loop output (not full context)

Performance Baseline Reference

Error Handling

Resources

• Lindy Prompt Guide
• Agent Steps
• Lindy Documentation

Next Steps

Proceed to lindy-cost-tuning for budget optimization.

Source: jeremylongshore/claude-code-plugins-plus-skills → skills/.curated/lindy-performance-tuning/SKILL.md

Also appears in: jeremylongshore/claude-code-plugins-plus-skills/plugins/saas-packs/lindy-pack/skills/lindy-performance-tuning/SKILL.md