🤖 AI Agents: Models, Tools, & The Secret To Making Them Work (Part 2)

Hi there!

In Part 1, we prepared all the "ingredients" to build an AI Agent. We now have:

1. Models (The Brains): AIs like GPT-4 or Claude 3.

2. Tools (The Hands): APIs, search functions, etc.

3. Evals (The Quality Check): A way to make sure the quality is good.

Now, in Part 2, we are going to learn the 4 "recipes" (Design Patterns) to "cook" those ingredients.

For each recipe, I will show you 2 ways:

• The Normal Prompt: How most people do it (and get average results).

• The Pro-level Prompt: How experts do it to get much better results (this is the real value).

Let's start with the first recipe.

I. Recipe 1: Reflection

This is the most powerful method you can start using right now without any special tools.

1. What Is The Idea?

It's simple: You ask the AI to review and critique its own work. Instead of accepting the first draft, you ask it to create a second draft.

2. The Normal Prompt (How Most People Do It)

This works, but the result is just okay.

• Step 1: The Draft Prompt

"Write a short blog post (100 words) about the benefits of drinking water."

• Step 2: The Reflection Prompt

"Is that post good? Make it better."

• The Problem: "Make it better" is too vague. The AI doesn't know what "better" means.

3. The Pro-Level Prompt (The Better Way)

We will use a specific "Rubric" (a list of rules). We force the AI to think in a structured way.

• Step 1: The Draft Prompt

"Write a short blog post (about 100 words) about the benefits of drinking water."

• The Result (Draft 1):

• Step 2: The Pro-level Reflection Prompt (Using a Rubric)

"You are a Professional Editor. Look at the <Draft> below. 
Now, grade that draft very strictly using this <Rubric>:
Opener (1-5 points): Is the first sentence boring? (1=very boring, 5=very interesting)
Specific Benefits (1-5 points): Are the benefits too general? ('good for skin' is general).
Call to Action (1-5 points): Is 'Drink more water' a useful tip?
After you are done grading, write a <Final_Version> that fixes all the problems you found."

• The Result (Draft 2 - Much Better):

💡 Pro-level Tip: Why does this work?

By using XML tags (like <Draft>) and a "Rubric," you force the AI to switch from "fast writing" mode to "structured analysis" mode. It must find the problems before it fixes them.

II. Recipe 2: Tool Use

This is how you give your Agent "hands" to interact with the real world.

1. What Is The Idea?

You give the AI access to outside tools, like a search API, your database, or your calendar. Modern models are trained to know when they need a tool.

2. The Normal Prompt (How Most People Do It)

This relies on the AI's "Zero-shot" (no examples) ability. It sometimes works, sometimes fails.

System Prompt:

"You are an assistant. You have a tool called get_weather(city). Use it if the user asks about weather."

The Problem: If the user asks a complex question ("Is it cold in Hanoi tomorrow? Should I bring a jacket?"), the AI might get confused or forget to use the tool.

3. The Pro-Level Prompt (The Better Way)

We will use "Guiding Examples" (Few-Shot Examples) to teach the AI exactly how we want it to think.

System Prompt:

"You are a smart assistant. You MUST follow the thought process below and use tools when needed.
# AVAILABLE TOOLS
search_product_database(product_name): Returns the ID and price of a product.
check_inventory(product_ID): Returns the stock number (how many are left).
# THOUGHT PROCESS AND EXAMPLES
You must always think step-by-step before you act.
Example 1:
User asks: "Do you sell blue t-shirts?"
Thought: "The user wants to know if we have a product. I need to search the database. I will use the search_product_database tool."
Action: search_product_database(product_name="blue t-shirt")
Example 2:
User asks: "How much is it and is it in stock?"
Thought: "This is a 2-part question. But I don't know what 'it' is. I need to ask the user for the product ID before I can check inventory."
Action: "Can you please tell me the product ID (SKU) you are asking about?"
Example 3:
User asks: "How many S-123 shirts are left?"
Thought: "The user gave me the ID (S-123) and wants the stock number. I will use the check_inventory tool."
Action: check_inventory(product_ID="S-123")
# START OF CONVERSATION"

💡 Pro-level Tip: Why does this work?

By giving "Guiding Examples," you don't just tell the AI what tools it has. You teach it how to handle situations (like when to use a tool, and more importantly, when to ask for more information). This reduces errors a lot. Tools like LangChain help manage this complex process.

III. Recipe 3: Planning

This is where we get into "Highly Autonomous" (thinks for itself) territory.

1. What Is The Idea?

You don't tell the AI the steps to do. You just give it a complex final goal and a list of tools. The Agent must then write a plan for itself to reach that goal.

2. The Normal Prompt (How Most People Do It)

"Goal: Plan a 3-day trip to New York for me.
Tools: search_flights, Google Hotels."

The Problem: The AI might make a very general plan, or it might start doing steps without thinking (e.g., it looks for hotels before it knows your budget).

3. The Pro-Level Prompt (The Better Way)

We will force the AI to use a "Plan and Critique" (Plan and Review) process before it takes any action.

System Prompt:

"You are an Expert Travel Planning Agent.
TOOLS: search_flights, Google Hotels, search_activities, ask_user_for_info
MANDATORY PROCESS:
When you get a goal, do not act immediately. You MUST follow this 4-step structure:
Thought: What is the user's final goal?
Initial Plan: Write down a step-by-step plan (1, 2, 3...) to reach the goal.
Self-Critique: Look at your plan. What information are you missing? (Example: Budget? Dates? Interests?). Where could this plan fail?
Final Plan: Write a new, better plan. The FIRST step of this new plan should ALMOST ALWAYS be to use ask_user_for_info to get all the missing information you found in step 3.
Only after you have shown these 4 parts can you start step 1 of the Final Plan."

• Now, when you give it a vague goal:

"Plan a trip to New York for me."

• The Agent's (Much Better) Response:

💡 Pro-level Tip: Why does this work?

This "Plan-and-Solve" (or "ReAct") method stops the AI from acting too fast. The "Self-Critique" step is the key. It forces the AI to think about what it doesn't know before it pretends to know.

IV. Recipe 4: Multi-Agent Workflow (Teamwork)

This is the most advanced design. Instead of one Agent doing everything, you create a "team" of special AI Agents.

1. What Is The Idea?

You create different "Roles" (Personas) for different AIs and have them "talk" to each other. For example: one Agent for research, one Agent for writing, and one Agent for critiquing.

2. The Normal Prompt (How Most People Do It)

Prompt 1: "Hey AI, find me 5 facts about blue whales."
Prompt 2 (copy-paste the result): "Hey AI, write a poem using these facts."

The Problem: You have to do the "glue work" (copy-paste). The Agents don't know about each other, and context is lost.

3. The Pro-Level Prompt (The Better Way)

We set up very detailed "Roles" and clear "Handoff Rules" (rules for passing the work).

• Goal: Create an Instagram ad campaign.

• Agent 1: The Data Analyst

"You are 'Data-Dave'.
Role: You are an analyst. You are COLD, logical, and ONLY speak in DATA.
Tools: search_trends, analyze_demographics.
Task: Analyze Instagram trends for 'sustainable running shoes'.
FINAL PRODUCT (Important): Your output MUST be a JSON object. Do NOT write 'hello' or explain.
Example output format:
{ "target_audience": "25-35, cares about environment", "key_trends": ["#sustainable", "#veganfashion"], "pain_point": "Sustainable shoes are often ugly" }"

• Agent 2: The Creative Writer

"You are 'Creative-Carla'.
Role: You are EMOTIONAL, creative, and you turn boring data into stories.
Task: You will get a JSON object from 'Data-Dave'. Do NOT repeat the data. Use it as inspiration to write 3 Instagram ads (50 words each) that solve the 'pain_point' Dave found.
FINAL PRODUCT: Only the 3 ad options."

• Agent 3: The Manager / Critic

"You are 'Manager-Mike'.
Role: You are the boss. You are practical and focused on results.
Task: You will get two things: (1) The original JSON from 'Data-Dave' and (2) The 3 ad options from 'Creative-Carla'.
Your Process:
Check if Carla's 3 ads actually solve the 'pain_point' from Dave's JSON.
Pick the 1 BEST ad.
Give 1 short reason why you picked it."
"You are 'Manager-Mike'.

💡 Pro-level Tip: Why does this work?

By creating specific personas and strict input/output formats (like JSON), you create an AI "assembly line." "Data-Dave" cannot be creative, and "Creative-Carla" cannot analyze data. This specialization creates a much better final product than one "general" Agent trying to do both.

V. Conclusion And Your Next Step

So, we have gone through all 4 recipes in Part 2:

1. Reflection: Ask the AI to critique itself using a rubric.

2. Tool Use: Teach the AI with guiding examples (few-shot).

3. Planning: Force the AI to "Self-Critique" its plan before acting.

4. Multi-Agent: Create a "team" of specialist Agents instead of one general Agent.

My advice: Don't try to build a complex Multi-Agent system right away.

Start simple.

Your homework: The next time you use AI, try the Pro-level Recipe 1: Reflection.

1. Ask it to do something (e.g., "write 3 email subject lines").

2. After it answers, don't stop. Use a "Rubric."

3. Say: "Now, grade those 3 subjects on a scale of 1-5 for: (1) Curiosity, (2) Urgency. Then, write 3 new subject lines that would score 5/5."

Congratulations! You just used an advanced AI Agent technique.

Would you like me to help you think of an idea for a simple "Tool Use" Agent you could build yourself?

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• Forget "Vibe Coding"! THIS Is The Future Of AI Building*

• This "Ugly" Banner Ad Formula Converts 100x Better*

• You're Using <10% Of ChatGPT's Power (Part 1)

• You're Using <10% Of ChatGPT's Power (Part 2)

• AI Superpowers For Your Work: The Top Tools That Actually Save Hours!

  *indicates a premium content, if any