Make AI Agents are FINALLY Here! (Full Tutorial)

Make.com AI Agents: A Deep Dive

Key Concepts:

• AI Agents: Language models with chat history and access to tools.
• Tools: Individual scenarios within Make.com that the AI agent can call.
• System Prompt: The core instructions and operating procedure for the AI agent.
• Scenario Inputs/Outputs: Data passed into and returned from tool scenarios.
• Thread ID: Used to maintain message history (though later clarified as potentially unnecessary with chat ID).
• Recursion Limit: The maximum number of actions an agent can perform based on one message.
• Agent Safeguard: A two-step process to ensure human approval before sensitive actions.

1. Overview of Make.com AI Agents

• Make.com has released AI agents, enabling automations with language models like OpenAI or Claude.
• These agents can access tools (represented as individual Make.com scenarios) to perform actions.
• The example showcased is a social media bot that researches topics, generates images, and posts to Facebook via Telegram.
• The agent receives a message via Telegram, decides which tool to call, and executes the corresponding scenario.

Example:

• A user sends "Post to Facebook please" via Telegram.
• The AI agent determines it needs to research the topic.
• It calls the "perplexity" tool scenario.
• The "perplexity" scenario researches the topic and returns the result.
• The AI agent then drafts a Facebook post based on the research.

2. Technical Setup of AI Agents and Tools

• Creating an Agent:
  1. Go to "AI Agents" in the left-hand menu and click "Create Agent."
  2. Add a connection to a language model (OpenAI, Claude, etc.).
  3. Enter an agent name.
  4. Select a model (e.g., Claude 3.7 Sonnet).
  5. Define the system prompt (core purpose and behavior).
• System Prompt Example:
  • "You are a Facebook posting AI agent designed to assist users with creating and posting content... Generate draft text... Ask for feedback... Ask if they want to generate an image... Ask for feedback on the image... Ask if they want to post anything else."
  • Includes instructions on tone of voice, understanding user intent, and calling specific tools.
• Creating Tools (Scenarios):
  1. Each tool is an individual scenario triggered "On Demand."
  2. Define "Scenario Inputs" (data the agent passes in, e.g., "prompt" for the perplexity tool).
  3. Define "Scenario Outputs" (data the tool returns, e.g., "research result text").
  4. Use the "Scenarios return output" module to pass the output back to the agent.
• Tool Examples:
  • Perplexity: Input: "prompt" (text). Output: "research result text" (text).
  • AI Image: Input: "image prompt" (text). Output: "image URL" (text). Uses file.AI service.
  • Facebook Posting: Input: "text" (text), "image URL" (text). Uses a router to handle text-only or image posts.
• Adding Tools to the Agent:
  1. In the agent settings, go to the "System Tools" section.
  2. Add the scenarios by name and provide a description for each.
  3. Ensure scenarios are active.

3. Demonstration and Workflow

• The demonstration shows the agent researching a topic ("AI agents"), generating an image, and posting to Facebook.
• The agent successfully calls the perplexity tool to research the topic.
• It then generates an image using the AI image tool.
• Finally, it posts the text and image to Facebook.
• The process involves multiple steps and interactions between the agent and the tool scenarios.

4. Managing Message History and Threads

• The "iterations from history count" setting in agent settings controls how many past messages are considered.
• A recursion limit of 300 is mentioned, potentially limiting the number of actions per message.
• The presenter initially used data stores to maintain thread IDs but later suggests that passing the chat ID from Telegram might be sufficient.

5. First Impressions and Shortcomings

• The agents work well in initial tests but are considered counterintuitive.
• Working with scenario inputs and outputs is awkward compared to N8N's native agents.
• Lack of real-time feedback on what the agent is doing is a significant drawback.
• Troubleshooting requires digging into the execution steps to see what happened.
• The tool-calling architecture is standard and makes sense for those with intermediate Make.com knowledge.

Quote:

• "I really expected make.com to be getting this right for beginners but really I think it's probably even easier to create these agents within N8N."

6. Advanced Features and Considerations

• Multi-agent systems (calling one agent from another) are possible, with synchronous and asynchronous execution options.
• Asynchronous execution uses webhooks for long-running processes.
• No built-in human-in-the-loop feature is available, increasing risk.
• The presenter recommends using an "agent safeguard" (a two-step approval process) for sensitive actions.

7. Agent Safeguard

• The agent safeguard is a two-step process that guarantees human approval before the agent posts to sensitive accounts.
• This involves a confirmation step where the user reviews and approves the content before it is posted.

8. Conclusion

• Make.com AI agents offer powerful automation capabilities but have usability challenges.
• The tool-calling architecture is well-structured, but the lack of real-time feedback and the complexity of scenario inputs make troubleshooting difficult.
• Implementing safeguards and carefully managing recursion limits are crucial for responsible agent usage.
• Despite the shortcomings, the agents have potential for automating complex workflows.