Quick Start Guide- Build Agent for Task Optimization

Introduction

Learn how to quickly run, debug, and optimize agents for web data extraction and e-commerce automation. Includes step-by-step instructions, examples like Amazon scraping and multi-platform price comparison, and model tuning strategies using GPT-4

Detail

Quick Start Guide

Core Principle: For an easy start, run tasks first, then optimize based on the gap between results and expectations.

Step 1: Direct Start

Go to the Run module, enter your user prompt, and start the Agent
If you find the Agent's task results differ from your expectations, proceed to the next step

Step 2: Identify Problems (Define the Gap)

Problem Type	Typical Symptoms	Specific Examples
Step Errors/Redundancy	Invalid clicks, repetitive operations, retry loops	Repeatedly clicking the same product on Amazon listing pages, getting stuck in endless loops
Extraction Errors	Missing data, incorrect data	Required product "name/price/rating" fields missing or content clearly abnormal
Output Format Errors	Wrong file format, incorrect order	Need "name/price/rating" order but CSV exports as "price/name/rating"

Step 3: Optimize System Instructions (Condition-Requirement-Constraint)

Instruction Structure:

Condition (When is this triggered?)
1. Define the scenario clearly, e.g., "When accessing Amazon..."
Requirement (How to solve it?)
1. Clearly define steps and output, e.g., "①Load all products→②Extract name/price→③Output CSV file"
Constraint (Optional - What absolutely cannot be done?)
1. Define limitations and solutions, e.g., "Prohibited from clicking ads."

Step 4: Complete Practical Examples

Example 1: Amazon Product Data Extraction Agent

Initial Run (No System Instructions)

User Task Instructions:

Search Amazon for "wireless headphones" category, extract the first 10 products' names, prices, ratings, and links, then export as CSV file.

Problems Discovered:

Missing price data (extraction error)
Incorrect links (URL error)

Optimized User Task Instructions

Objective: Search Amazon for "wireless headphones", extract complete information for the first 10 products, complete within 30 steps.

Specific Requirements:
1. Search keyword: "wireless headphones"
2. Extract fields: Product Name, Price, Rating, Product Link
3. Product quantity: First 10 products
4. Output format: CSV file, column order as "Product Name,Price,Rating,Product Link"
5. Step limit: Complete within 30 steps
6. Data quality: Ensure each field has data or mark as "data missing"

Optimized System Instructions

You are a professional Amazon product data extraction specialist. You need to efficiently and accurately extract product information from Amazon website.

【Extraction Process Optimization】
When on Amazon product listing pages:
① Load complete listing page → ② Click each product card once in sequence → ③ Extract title/price/rating/link → ④ Return to previous page and continue until all specified products collected
Constraints: Prohibited from clicking advertisement positions, prohibited from repeatedly clicking same product

【Data Extraction Standards】
When entering Amazon product detail pages:
① Enter detail page → ② Read product title, current price, overall rating, product link in sequence → ③ If any field "name/price/rating/link" is missing, read page section by section and search once more; if still not found, mark as "data missing"
Constraints: Prohibited from estimating values to fill gaps; all data must reference actual page data

【Output Format Control】
When exporting all extraction results as CSV file:
① Compile all records → ② Generate CSV file with header row first, column order strictly as "Product Name,Price,Rating,Product Link" → ③ Ensure data integrity and format consistency
Constraints: Replace line breaks or extra commas in fields with spaces first, standardize price to numeric format

【Quality Assurance】
- Each product must include 4 fields: name, price, rating, link
- Price format: Keep only numbers and decimal points, remove currency symbols
- Rating format: X.X/5.0 or data missing
- Link format: Complete Amazon product URL

Example 2: Multi-Platform Price Comparison Agent

System Instructions

You are a professional e-commerce price comparison analyst. You need to search for the same product across multiple e-commerce platforms and conduct price and basic information comparison analysis.

【Platform Access Strategy】
When accessing different e-commerce platforms:
① Visit specified platforms in sequence → ② Search using same keywords → ③ Select most relevant product for data extraction → ④ Record platform name, product information, and price
Constraints: Select only 1 best matching product per platform, avoid staying too long on single platform

【Data Standardization】
When extracting data from different platforms:
① Standardize product name format (remove platform-specific identifiers) → ② Standardize price format (unified currency) → ③ Record platform-specific information (shipping methods, promotional info)
Constraints: Prices must be converted to numeric format, platform names use standard abbreviations

【Comparative Analysis】
When all platform data collection is complete:
① Organize all platform data → ② Calculate price differences and percentages → ③ Mark lowest and highest price platforms → ④ Generate comparison report
Constraints: Price comparisons must be based on same or similar products, note product differences

User Task Instructions

Task: Compare "iPhone 15 Pro 128GB" price information across Amazon, eBay, and Best Buy.

Execution Steps:
1. Visit Amazon, eBay, and Best Buy respectively
2. Search for "iPhone 15 Pro 128GB"
3. Select most matching product and extract information
4. Record: Platform name, product title, price, seller name, shipping info
5. Generate price comparison table
6. Mark best choice and price difference percentages

Output with csv file
Step Limit: Complete within 45 steps

Step 5: Model and Temperature Adjustment

Task Type	Recommended Model	Temperature Range	Reason
Data Extraction (Consistency)	GPT-4.1-Mini	0~0.3	Low randomness, ensures data accuracy
Multi-site Complex Parsing	GPT-4.1	0.3~0.6	Moderate flexibility, handles page variations
Analysis Report Generation	GPT-4.1	0.6~0.8	Higher creativity, generates deep analysis
High Uncertainty Scenarios	GPT-4.1	0.7~1	High randomness, improves diversity

Model Selection Guidelines:

Use GPT-4.1-Mini (10 credits/step):

Simple, well-defined tasks
Standardized data extraction
Need quick, simple answers

Use GPT-4.1 (30 credits/step):

Complex, uncertain tasks
Multiple sites with different page structures
Need to handle captchas or infinite scroll

Temperature Settings:

Low (0-0.3): Specific information searches, consistent data extraction
Medium (0.4-0.6): Balanced tasks requiring some flexibility
High (0.7-1.0): Creative tasks like content generation, diverse outputs needed

Step 6: Complete Optimization Process Example

Full Optimization Workflow

Initial State:

System Instructions: Empty
User Task: Search Amazon for "laptop computers" first 5 products

First Run Results:

Issues: Repeated clicks, missing prices, format chaos
Steps: 45
Success Rate: 60%

First Optimization: Added step control system instructions:

When on Amazon product listing pages:
① Load listing page → ② Click each product once in sequence → ③ Extract info then return → ④ Avoid repetitive operations

Second Run Results:

Issues: Steps reduced, but data quality is still problematic
Steps: 32
Success Rate: 75%

Second Optimization: Added data quality control:

When extracting product information:
① Product name: Complete title → ② Price: Numeric format only → ③ Rating: X.X format → ④ Mark missing data as "N/A"

Third Run Results:

Issues: Data quality improved, but output format still needs adjustment
Steps: 28
Success Rate: 90%

Final Optimization: Added output format control:

When exporting CSV:
① Header: Product Name,Price,Rating,Link → ② Data cleaning: Remove special characters → ③ Format validation: Ensure 4 fields per row

Final Results:

Steps: 25 (44% reduction)
Success Rate: 95%
Output: Complete CSV file with proper formatting

Advanced Configuration Examples

E-commerce Price Tracker Agent

System Instructions: You are an e-commerce price tracking specialist...
Temperature: 0.2 (Low - for consistent data extraction)
Model: GPT-4.1-Mini (Cost-effective for structured tasks)

Content Research Agent

System Instructions: You are a content research analyst...
Temperature: 0.6 (Medium - for balanced analysis)
Model: GPT-4.1 (Better for complex reasoning)

Creative Content Generator

System Instructions: You are a creative content strategist...
Temperature: 0.8 (High - for diverse creative outputs)
Model: GPT-4.1 (Better for nuanced creativity)

Best Practices Summary

System Instructions Writing Tips:

Layered Structure: Group by functional modules (extraction, processing, output)
Specific Constraints: Clearly define what can and cannot be done
Exception Handling: Define how to handle missing data, page errors
Quality Standards: Set data format and completeness requirements

User Task Instructions Essentials:

Clear Objectives: Specific quantities, scope, format requirements
Step Limits: Reasonable step budgets
Quality Requirements: Data accuracy and completeness standards
Output Specifications: File format, naming, structure requirements

Common Pitfalls to Avoid:

Overcomplicating initial instructions
Adding constraints without testing first
Ignoring cost vs. performance trade-offs
Not specifying output requirements clearly
Not handling edge cases (missing data, format variations)

Performance Optimization Tips:

Start Simple: Begin with basic functionality, add complexity gradually
Test Iteratively: Run → Identify issues → Optimize → Repeat
Monitor Costs: Balance model choice with budget constraints
Validate Outputs: Ensure data quality meets requirements
Document Changes: Track what works and what doesn't

Remember: The goal is to systematically bridge the gap between expectation and reality through optimization, not to create perfect instructions from the start.

Related reading: For practical examples, implementation details, and trade-offs, continue with Build an AI Web Research Agent with Primary Sources.