Make.com Just Released AI Agents–Learn Everything in 30 Mins

Key Concepts

AI Agent, make.com, n8n, Telegram, OpenAI, Chat GPT, GPT-4, Large Language Model (LLM), Webhook, BotFather, API Token, Thread ID, System Instructions, Sub Agent, Scenario Inputs/Outputs, Array Aggregator, Collection, Blueprint, RAG System, Voice AI Agent, Slack AI Agent, Max Output Tokens, Recursion Limit, Iteration History Count.

AI Agent Demo and Comparison

The video demonstrates make.com's newly launched AI agent, positioning it as a competitor to n8n's existing AI agent. The presenter suspects make.com's agent might be slightly behind n8n's currently but will likely catch up quickly. The core functionality involves receiving messages (via Telegram in this example) and using AI to determine the appropriate action and tool to execute.

• Example 1: Sending an Invoice: The user sends a Telegram message "Hey can you please send an invoice over to Jonno?". The AI agent transcribes the message, identifies the "send email" tool, and sends an invoice to the specified recipient. The invoice document is processed almost immediately.
• Example 2: Sending an Email: The user asks the AI agent to email Jonno, informing him that the invoice has been sent and requesting a review. The AI agent uses the Gmail module to look up Jonno's email address and sends the email.
• Example 3: Checking Calendar: The user asks the AI agent to check their Google Calendar for tomorrow. The AI agent retrieves the calendar events and sends them back to the user.

Blueprint Availability and Use Cases

The presenter offers free blueprints for the demonstrated AI agent and other use cases, downloadable from the video description. These blueprints can be imported into make.com for immediate use.

• Blueprint Examples: Scraping data from the internet (Google, Facebook, Instagram), building a RAG (Retrieval-Augmented Generation) system for customer messaging, creating a personal AI assistant, developing a voice AI agent, and building a Slack agent.
• Community Access: The presenter promotes a community where users can access more blueprints, participate in weekly calls, and learn about make.com and related technologies.

Building an AI Agent from Scratch

The video provides a step-by-step guide to building an AI agent in make.com.

1. Setting up the Main AI Agent

• Creating a New Scenario: Start by creating a new scenario in make.com.
• Selecting the AI Agent Module: Choose the "AI Agent" module.
• Creating a New AI Agent: Create a new agent, selecting the desired Large Language Model (LLM) (e.g., Chat GPT, GPT-4). The presenter uses GPT-4 Mini.
• Naming and Prompting the Agent: Name the agent (e.g., "Jonno's AI Assistant") and provide a prompt defining its role (e.g., "You're an AI assistant for Jonno helping him complete his tasks").
• Configuration Settings:
  • Max Output Tokens: Limits the number of tokens used per run to prevent excessive usage.
  • Recursion Limit: Sets a limit on the number of recursive tool calls to prevent infinite loops.
  • Iteration History Count: Determines how many past messages the AI agent considers when formulating a response (default: 10).
• Setting up the Telegram Trigger:
  • Adding the Telegram Module: Add the "Telegram" module and select "Watch Updates" to listen for incoming messages.
  • Creating a Webhook: Create a webhook in Telegram using BotFather to connect the Telegram bot to make.com.
    • Search for "BotFather" in Telegram.
    • Type /newbot to create a new bot.
    • Provide a name and username for the bot (username must end with "bot" or "_bot").
    • Copy the API token provided by BotFather.
  • Connecting Telegram to make.com: Paste the API token into the Telegram module in make.com.
• Sending Messages Back to Telegram:
  • Adding the "Send a Text Message or Reply" Module: Add the Telegram module to send messages back to the user.
  • Mapping the Chat ID: Map the "chat ID" from the incoming Telegram message to the "chat ID" field in the send message module.
  • Mapping the AI Agent Response: Map the AI agent's response to the "text" field in the send message module.
• Thread ID Importance: The thread ID is crucial for maintaining conversation history. It ties the conversation to a specific Telegram chat, allowing the AI agent to remember previous messages.
• System Instructions: Provide additional context to the AI agent to guide its behavior and tool usage.
• Message Mapping: Map the incoming Telegram message to the AI agent's message input.

2. Creating a Sub Agent (Tool)

• Creating a New Scenario: Create a new scenario for the sub agent (e.g., "Search Calendar").
• Selecting the Google Calendar Module: Choose the "Google Calendar" module and select "List Events".
• Scenario Inputs: Define scenario inputs for passing data from the main AI agent to the sub agent (e.g., "start time", "end time"). These inputs are used to specify the date range for searching calendar events.
• Mapping Inputs to Calendar Module: Map the scenario inputs to the corresponding fields in the Google Calendar module.
• On-Demand Trigger: Ensure the sub agent's trigger is set to "On Demand" or "Instant Trigger". This is essential for AI agents to function correctly.
• Activating the Scenario: Activate the sub agent scenario.
• Adding the Tool to the Main AI Agent:
  • In the main AI agent, add a new tool and paste the name of the sub agent scenario.
  • Provide a description for the tool, explaining when the AI agent should use it (e.g., "Call this scenario when the user is looking to find events on their calendar").
• Tool Classification and System Prompts:
  • Classify the tool (e.g., "search calendar") and provide additional information (e.g., "find calendar events for Jonno").
  • Include rules and fallback values (e.g., if no end time is provided, create an end time one day after the start time).

3. Handling Outputs (If Necessary)

• Array Aggregator: If the sub agent returns multiple events, use an array aggregator to merge them into a list.
• Scenario Outputs: Define scenario outputs to structure the data being sent back to the main AI agent.
  • Model the output structure to match the data being returned (e.g., a list of collections, where each collection represents a calendar event).
• Mapping the Array to the Output: Map the aggregated array to the scenario output.

Addressing Date Issues

The presenter encountered an issue where the AI agent was pulling in dates from the past (2023) instead of the current date. To resolve this, the presenter added a rule to the main AI agent's system prompt specifying today's date using the now function.

Bug and Workaround

The presenter noted a bug where the description for the tool in the main AI agent kept getting deleted. As a workaround, the presenter emphasized the importance of ensuring the description is programmed in for the tool to function correctly.

Conclusion

The video concludes by highlighting the potential of make.com's AI agent and acknowledging that it is still in beta and may have some bugs. The presenter expresses a desire for the ability to have multiple tools within a single workflow, rather than always having to create sub agents. The presenter encourages viewers to join their community for additional help, blueprints, and live calls. The main takeaways are the power and flexibility of make.com's new AI agent, the importance of proper setup and configuration, and the potential for automating a wide range of tasks.