How I’d Teach a 10 Year Old to Build Agentic Workflows (Claude Code)

Understanding Agentic Workflows

Introduction to Agentic Workflows

• Agentic workflows revolutionize AI automation by enabling self-fixing processes that are significantly faster than traditional methods, requiring no coding.

• The video aims to clarify agentic workflows for beginners, ensuring even a child can grasp the concept.

Traditional vs. Agentic Workflows

• Traditional workflow automation involves detailed step-by-step instructions where users map out every action and data flow.

• In contrast, agentic workflows allow users to specify desired outcomes without detailing the steps; the system autonomously determines how to achieve those results.

Practical Analogy

• The analogy of ordering food illustrates the difference: traditional automation is like cooking from a recipe, while agentic workflows resemble simply requesting a meal at a restaurant.

• Users provide goals and may answer clarifying questions, allowing the system to tailor its approach based on user preferences.

Example Use Case

• For instance, if someone needs contact information for dentists in Chicago, they only need to state their requirement; the system will handle data scraping and email personalization autonomously.

Visualizing Workflow Differences

• A side-by-side comparison shows traditional automation as complex with multiple steps versus agentic workflows being straightforward with minimal input leading directly to output.

Building an Agentic Workflow

Getting Started with Tools

• Users are guided through setting up their environment using Visual Studio Code (VS Code), which is free and essential for working with Claude Code.

Installing Necessary Extensions

• Instructions include searching for and installing the Claude Code extension within VS Code after signing up for a paid plan starting at $17/month.

Project Setup in VS Code

• Users must create or open a project folder in VS Code before proceeding with building their agentic workflow.

Introduction to Claude Code and WAT Framework

Overview of the Environment

• The speaker introduces a completely blank project in Claude Code, highlighting the file explorer interface where workflows will be built.

• The left side displays files related to the "Agentic Workflows demo," while the right side features a chat interface for interaction with the coding assistant.

Understanding WAT Framework

• The framework used is called WAT, which stands for Workflows, Agents, and Tools. This structure is essential for building agentic workflows.

• Workflows are represented in markdown format, making them easy to read with natural language elements like headers and bullet points. Tools are implemented using Python code.

Distinction Between Workflows and Tools

• A metaphor comparing workflows to recipes (instructions) and tools to ingredients (actions), illustrating how an agent (chef) uses both to create outcomes (cakes or workflows).

• Emphasizes that workflows guide actions while tools execute them; understanding this distinction is crucial for effective workflow creation.

Setting Up Claude Code

Importance of Cloud.md File

• To initiate a new project in Claude Code, a cloud.md file must be created. This file serves as job instructions for the AI agent.

• The cloud.md file outlines how Claude should operate within the WAT framework, detailing three layers: workflows (instructions), agents (decision makers), and tools (executions).

Operational Guidelines

• Discusses why structured frameworks matter—accuracy diminishes rapidly without clear guidelines; each step's accuracy impacts overall success rates significantly.

• Introduces operational strategies such as seeking existing tools first, adapting when failures occur, and maintaining current workflows.

File Structure Management

Organizing Files Effectively

• Highlights the necessity of organizing files within Claude Code to prevent confusion; proper structuring aids both user comprehension and AI functionality.

• Mentions different modes available when interacting with Cloud Code: ask before edits, edit automatically, plan mode, and bypass permissions—each affecting AI autonomy levels during operation.

How to Use Claude for Project Setup and Workflow Automation

Overview of Claude's Functionality

• Claude can be configured to edit automatically, request edits, or bypass permissions entirely. Users must enable the "allow dangerously skip permissions" option in settings for full control.

• The speaker demonstrates using Claude in bypass permissions mode to set up a project by reading a provided claude.md file.

Project Initialization Process

• After reading the claude.md file, Claude identifies existing project requirements and structures. It creates an organized folder structure including temporary folders and tools.

• Users can observe real-time updates as Claude completes tasks like creating a to-do list and setting up necessary files for the project.

Building Workflows with Remote Job Applications

• The speaker discusses applying for 622 remote jobs found online, emphasizing the difficulty of manual logging due to multiple pages of listings.

• To streamline this process, they plan to use Firecrawl, a tool that scrapes website data and compiles it into an Excel sheet.

Utilizing Firecrawl with MCP Framework

• Firecrawl allows various actions such as scraping data from websites. The speaker highlights its capability to convert web content into LLM-ready data.

• The Model Context Protocol (MCP) is introduced as a framework that enables agents like Claude to utilize multiple tools efficiently without user intervention.

Practical Application of MCP in Workflow Automation

• An analogy is drawn comparing MCP's functionality to grocery shopping where an agent retrieves all needed ingredients from one source rather than visiting multiple stores.

• The speaker explains how MCP simplifies task execution by allowing agents access to various tools through a single interface.

Installation of Firecrawl MCP Server

• Instructions are provided on installing the Firecrawl MCP server within cloud code. Users need their API key but should avoid storing it in conversation history for security reasons.

How to Securely Initialize a Project with API Keys

Setting Up the Environment

• The speaker emphasizes the importance of securely initializing a project by using an .env file for storing sensitive information like API keys instead of hardcoding them directly into the code.

• An API key placeholder is added to the .env file, prompting the user to input their actual API key. The speaker mentions signing up for Firecrawl to obtain a free API key.

Adding and Saving the API Key

• The speaker copies their Firecrawl API key and pastes it into the .env file, ensuring it's saved properly in the project configuration.

• After adding the API key, they request assistance from Claude (an AI tool) to run a command that installs the Firecrawl MCP server.

Security Considerations

• The speaker discusses potential risks associated with exposing an API key during command execution but reassures that since this is a free key, it poses minimal risk. They recommend rotating keys after use for better security practices.

• Best practices are highlighted: always store sensitive keys in an .env file to prevent exposure when pushing code to public repositories.

Building an Agentic Workflow

• Transitioning into building an agentic workflow, they switch to "plan mode" and prepare to scrape job opportunities from a provided URL containing 622 listings across multiple pages.

Interacting with Claude for Workflow Development

• The speaker instructs Claude on scraping job listings and organizing them into an Excel sheet while encouraging questions for clarity on project requirements.

• Claude begins formulating questions regarding additional details needed for scraping job data, such as whether only listing titles or full descriptions should be included.

Reviewing Plans and Execution

• After submitting initial parameters, Claude presents its plan which includes creating tools and workflows necessary for executing the scrape task effectively.

• A review of Claude's comprehensive plan reveals it will create specific tools and workflows aimed at efficiently gathering job listing data.

Finalizing Workflow Execution

• Upon approval of Claude's plan, execution begins successfully resulting in 209 job postings being scraped. Metrics are displayed post-execution indicating successful completion of tasks.

Scraping Job Listings and Managing Context

Overview of the Scraping Tool

• The speaker introduces a new tool for scraping daily remote job listings, enhancing its functionality through improved workflows. This allows for better performance in future scraping tasks.

• A temporary file is created to store the scraped data, which includes various job details such as title, type, location, experience level, salary, and more. The output consists of 209 total job postings.

Context Management

• The concept of "context rot" is discussed; it refers to the degradation of conversation quality with an AI model as context accumulates beyond a certain limit (60%). To mitigate this, the speaker compacts the workflow and conversation history while retaining essential information.

Expanding Job Search Parameters

• The speaker attempts to scrape jobs from a broader dataset without filters but realizes that only 52 sales jobs are available in Europe out of 409 total sales jobs found globally. This prompts a reevaluation of search parameters to meet their goal of obtaining 500 listings.

• The AI suggests expanding the search criteria by including US-based sales jobs or broadening to all job types when faced with limitations in available data. This demonstrates adaptive problem-solving capabilities within the tool's framework.

Final Output and Data Validation

• After adjusting parameters, the tool successfully finds 372 sales jobs across Europe and the US, presenting them in an organized Excel sheet with pre-made filters for easy navigation through job titles and other relevant information.

Testing New Requests

• A hypothetical scenario is presented where the speaker requests assistance in generating leads for dentists in the United States using an Excel format. The AI checks available tools and APIs before proceeding with lead generation tasks, showcasing its ability to adapt to diverse requests beyond initial programming constraints.

Scraping Dentist Leads: A New Workflow

Overview of the Scraping Process

• The speaker discusses creating a tool similar to an existing job scraper, specifically for scraping dentist leads from directories.

• Initial attempts yield only two dentists found due to parsing issues; adjustments are made to the regex pattern for better results.

• After refining the tool, 120 unique dentist leads are successfully scraped from four major cities, showcasing the effectiveness of the new workflow.

Data Collected and Tool Functionality

• The output includes comprehensive data such as phone numbers, addresses, websites, specialties, and listing URLs formatted in an Excel sheet.

• The speaker emphasizes that despite limited initial information provided to the AI agent, it produced valuable results efficiently compared to manual efforts.

Advantages of Using AI Agents

• Multiple agents can be deployed simultaneously with different workflows; this allows testing various methods and selecting the most effective one based on results.

• The first workflow example is introduced: scraping job listings from daily remote.com. It highlights how clear objectives and required inputs streamline processes.

Common Mistakes in Defining Workflows

• A key mistake is being vague about goals; specific details about desired outcomes are crucial for effective agent performance (e.g., specifying lead types).

• Another common error is failing to define what "done" looks like; clear finish lines prevent overcomplication and inefficiency in task execution.

Best Practices for Effective Agentic Workflows

• Providing detailed project requirements helps agents understand tasks better—similar to giving specifications to human developers.

• Clear input-output expectations ensure consistent results; specificity in requests (e.g., exact number of profiles needed) enhances efficiency.

Gentic Workflows: Revolutionizing Automation

The Challenges of Traditional Workflow Automation

• Traditional workflow automation often leads to debugging loops where edge cases break the system, requiring extensive time spent on logs and error messages.

• Gentic Workflows simplifies this process by having an agent that automatically identifies issues and suggests solutions, effectively self-healing without user intervention.

Benefits of Gentic Workflows

• Users can multitask while the agent builds workflows, allowing for a more efficient use of time as it handles errors autonomously.

• Natural language control in Gentic Workflows eliminates the need for users to learn every node or API documentation; users simply explain their needs, and the system adapts accordingly.

• The agent learns from past issues and updates automations automatically, enhancing efficiency over time.

Human vs. Scheduled Triggered Automations

• There is a distinction between human-triggered events (e.g., writing proposals during calls) where the agent assists in real-time versus scheduled automations that require manual deployment by the user.

• Understanding this difference is crucial for effective use of Gentic Workflows, as only certain aspects are self-healing while others depend on user input.

Future of Automation with Gentic Workflows

• The landscape of automation is rapidly evolving towards fully autonomous systems capable of managing themselves and improving independently.

• Emphasizes the importance of adapting to new challenges in automation technology; resources are available for those looking to enhance their skills within a community setting.