AI Workflows Just Got 10x Easier (Goodbye Zapier & Make & n8n?)

Agentic Workflows: A Detailed Breakdown & Practical Implementation

Key Concepts:

• Artificial Intelligence (AI): Systems trained on vast datasets to respond to inputs.
• AI Automation: Utilizing tools (Make.com, Naden) to automate tasks previously done manually with AI.
• AI Agents: AI systems with a “brain” (reasoning ability) and “memory” (context retention), capable of independent action.
• Agentic Workflows: Workflows built and run by two “digital employees” – one to build the automation, the other to execute and maintain it.
• Orchestration: The process of a “manager” AI agent routing tasks to specialized “employee” AI agents.
• Self-Healing: An agentic workflow’s ability to identify and automatically fix errors in its own code.
• System Prompt: Instructions provided to an AI agent defining its role and tasks.

1. The Four Layers of AI

The video outlines a progression of AI capabilities, broken down into four layers:

• Layer 1: Artificial Intelligence: This is the foundational layer, exemplified by tools like ChatGPT. It operates on a simple input-output model. A user provides a prompt, and the AI generates a response based on its training data. The key limitation is its manual nature – requiring users to copy/paste information and manage the entire process.
• Layer 2: AI Automation: This layer introduces automation using platforms like Make.com and Naden. These tools connect AI with other applications, eliminating manual steps. For example, an email arriving in your inbox can automatically trigger an AI-generated response and send it back. However, these automations are brittle; any change in connected systems can break the workflow, requiring manual intervention.
• Layer 3: AI Agents: AI Agents add a layer of intelligence and memory to AI workflows. They function like digital employees, capable of reasoning and remembering past interactions. They consist of a “brain” (like ChatGPT’s reasoning engine) and a “memory” (allowing them to retain context). Crucially, they require clear instructions in the form of a “system prompt.” They can also be connected to various tools (Gmail, Google Sheets, etc.).
• Layer 4: Agentic Workflows: This is the most advanced layer. It utilizes two AI agents: one to build the automation workflow and another to run and maintain it. This eliminates the need for users to learn complex automation tools like Naden or Make.com.

2. The Problem with Traditional Automation & the Rise of Agentic Workflows

The speaker highlights the limitations of traditional AI automation: its fragility and the technical expertise required to build and maintain workflows. Agentic workflows address these issues by automating the creation and maintenance of automations.

Quote: “No longer do you have to come in here and learn these tools like Naden and Make.com because you can essentially get the results you're looking for just by sending off plain language like you're used to already inside ChatGPT.”

3. Orchestration: The Manager-Employee Analogy

The concept of “orchestration” is explained using a business management analogy. An AI agent acts as a manager, receiving tasks and delegating them to specialized “employee” agents.

• Example: A task to send an email is routed to a “Gmail employee” agent, while a task involving document creation would be sent to a “Google Drive employee.” The manager agent determines the best agent for each task. This ensures efficient task completion.

4. AI Agents vs. Agentic Workflows: The Self-Healing Distinction

The video draws a critical distinction between AI agents and agentic workflows, focusing on the concept of “self-healing.”

• AI Agents: An AI agent might repeatedly fix the same error, but doesn’t proactively address the root cause. It’s like an employee who consistently solves a recurring problem without identifying and eliminating its source.
• Agentic Workflows: An agentic workflow identifies the recurring error and modifies the underlying code to prevent it from happening again. This proactive approach is what sets agentic workflows apart.

Quote: “With an agentic workflow, you can think about it like an employee who says, ‘Hm, hold on a second. We're encountering the same mistake every single day. Why don't we just go ahead and change the process so this mistake never happens again?’”

5. Practical Implementation: Web Scraping with Anti-Gravity

The video demonstrates building an agentic workflow using Google’s Anti-Gravity platform. The example focuses on web scraping email addresses from coffee shop listings on Google Maps in Toronto.

• Step-by-Step Process:

  1. Download & Install Anti-Gravity: The desktop application is downloaded and installed.
  2. Create a New Folder: A new folder is created within Anti-Gravity to contain the workflow.
  3. Prompt the Agent: A natural language prompt is provided to the agent, outlining the desired workflow: scrape two coffee shops in Toronto, extract email addresses, save to a CSV file, and avoid opening a browser.
  4. Agent Generates Code: Anti-Gravity generates a Python script based on the prompt.
  5. User Approval: The agent requests permission to execute the code, requiring user approval.
  6. Workflow Execution & Iteration: The agent executes the script, scrapes data, and identifies/fixes formatting issues in the email extraction logic.
  7. Save as Reusable Task: The completed workflow is saved as a reusable task named “find local businesses and extract contact emails.”
  8. Run Reusable Task: The saved task can be run again with different parameters (e.g., different cities, business types) by simply prompting the agent.

• Technical Details: The workflow involves Python scripting, web scraping, and CSV file manipulation. Anti-Gravity handles the coding complexities, allowing users to interact with the system using natural language.

6. Anti-Gravity Specifics

• Three Panes: The Anti-Gravity interface consists of three panes: a file/folder management pane (like Google Drive), a central pane displaying file contents, and a right-hand pane housing the “junior developer” agent.
• Screenshot Input: Users can input prompts by screenshotting text, which Anti-Gravity can then read and execute.
• Probabilistic Nature: Anti-Gravity, being a large language model, produces slightly different results each time, but generally converges on the same solution.

7. Resources & Further Learning

The speaker promotes a “school community” offering three transformations:

• Transformation 1: Learning and applying AI/automation tools.
• Transformation 2: Building and running an AI automation agency.
• Transformation 3: Automating tasks within an existing business.

Conclusion:

The video effectively demystifies agentic workflows, positioning them as the next evolution in AI automation. By automating both the creation and maintenance of workflows, agentic workflows promise to empower users without requiring extensive technical expertise. The practical demonstration with Anti-Gravity highlights the potential of this technology to streamline complex tasks and unlock new levels of productivity. The key takeaway is that agentic workflows represent a shift towards more autonomous and self-healing AI systems.