AI Agents Intro

Simple Definition:

An AI Agent is an LLM-powered system that can reason about a task, decide what tools to use, execute actions, and observe results to accomplish a goal - all without constant human hand-holding.

Think of it like this: a regular chatbot is like a customer care executive who can only answer questions from a script. An AI agent is like a smart employee who can think on their feet, use different tools (computer, phone, email), and figure out the best way to solve your problem.

Real-World Analogy - Swiggy Order System:

Imagine you tell a Swiggy support agent: "My order is late, find out why and get me a refund."

• Regular Chatbot: "Sorry for the inconvenience. Please contact support." (just text response)
• AI Agent: Checks order status API -> Sees delivery partner stuck -> Checks refund policy -> Initiates refund -> Sends you confirmation. All autonomously!

The Core Agent Loop:

1. Perceive: Understand the user request and current context
2. Reason: Think about what needs to be done (this is where LLM shines)
3. Plan: Decide which tool/action to use next
4. Act: Execute the chosen tool or action
5. Observe: Look at the result and decide if goal is achieved
6. Repeat: If not done, go back to step 2 with new information

Agent vs Chatbot vs Workflow - Key Differences:

What is ReAct?

ReAct (Reasoning + Acting) is a paradigm where the LLM alternates between thinking (reasoning about the problem) and acting (using tools). It was introduced in a 2022 paper by Google and Princeton researchers.

The key insight: when you let an LLM "think out loud" before acting, it makes much better decisions about which tool to use and what input to provide.

ReAct Loop - Step by Step:

User: "What is the current stock price of Reliance?"

--- Iteration 1 ---
Thought: I need to find the current stock price of Reliance Industries.
Action: stock_price("RELIANCE.NSE")
Observation: Current price: Rs 2,847. 52-week high: Rs 3,024.

--- Iteration 2 ---
Thought: I have the price. Now I should check analyst recommendations.
Action: search("Reliance Industries stock analysis 2026")
Observation: Analysts suggest Reliance is fairly valued...

--- Iteration 3 ---
Thought: I now have both price data and analyst opinions.
Action: respond_to_user(final_answer)

Why ReAct Works So Well:

• Transparency: You can see reasoning at each step, making debugging easy
• Grounding: Actions provide real data that prevents hallucination
• Flexibility: Agent decides dynamically which tool to use
• Error Recovery: If a tool call fails, the agent can reason about alternatives

ReAct Limitations:

• Sequential Thinking: One step at a time - cannot plan ahead efficiently
• Can Get Stuck: Sometimes loops doing the same action repeatedly
• Token Cost: Each thought + action + observation consumes tokens
• No Global Plan: Does not create an upfront plan, just reacts step-by-step

What is Plan-and-Execute?

Instead of thinking one step at a time (like ReAct), Plan-and-Execute first creates a complete plan with all the steps needed, then executes each step one by one. If a step fails or new information changes things, it replans.

Think of it like this: ReAct is like navigating without a map, making decisions at each turn. Plan-and-Execute is like using Google Maps - you plan the full route first, then follow it, rerouting if there is a road block.

Example - Flipkart Price Comparison Agent:

User: "Compare iPhone 16 prices across Flipkart, Amazon, and Croma"

--- PLANNING PHASE ---
Plan:
  Step 1: Search Flipkart for iPhone 16 price and offers
  Step 2: Search Amazon for iPhone 16 price and offers
  Step 3: Search Croma for iPhone 16 price and offers
  Step 4: Check current bank offers on each platform
  Step 5: Compare all prices and recommend best deal

--- EXECUTION PHASE ---
Step 1: Rs 79,900 (MRP Rs 89,900), 10% HDFC discount
Step 2: Rs 78,499, 5% cashback on Amazon Pay
Step 3: Rs 81,990, exchange bonus Rs 5,000

--- REPLANNING ---
Exchange offers change the math. Updating plan...

Final Answer: Best deal is Amazon at Rs 78,499...

When to Use Which Pattern:

What is Tool Calling?

Tool calling (function calling) is the mechanism by which an LLM can request to execute external functions. Instead of hallucinating an answer, the agent says "I need to call this function with these parameters" and the system executes it.

Modern LLMs like GPT-4, Claude, and Gemini output structured tool calls as part of their response.

Types of Tools Agents Use:

• Information Retrieval: Web search, database queries, API calls, RAG retrieval
• Data Processing: Calculator, code interpreter, data transformation
• Communication: Send email, Slack message, push notification
• System Actions: Create file, deploy code, update database record
• Human-in-the-Loop: Ask user for clarification, get approval

Best Practices for Tool Design:

• Clear Names: search_products not sp
• Rich Descriptions: Explain when to use the tool and what it returns
• Typed Parameters: Use JSON Schema with types and constraints
• Helpful Errors: Return error messages the LLM can reason about
• Limit Count: 10-15 tools max per agent
• Idempotent: Tools should be safe to retry on failure

Key Implementation Details:

• System Prompt: Defines the agent persona, available tools, safety constraints, and output format. A bad system prompt = a bad agent.
• Message History: Full conversation (thoughts + actions + observations) passed to LLM each iteration. Token usage grows with each step.
• Max Iterations: Always set a limit (e.g., 10) to prevent infinite loops.
• Token Budget: Track usage per run. Set cost limits to prevent runaway spending.
• Timeout: Set per-tool and total timeouts.

Common Pitfalls:

• Infinite Loops: Agent keeps calling same tool. Solution: detect repetition, force termination.
• Hallucinated Tools: Agent invents non-existent tools. Solution: strong system prompt + validation.
• Context Overflow: Too many iterations fill context. Solution: summarize older turns.
• No Guardrails: Agent takes destructive actions. Solution: human-in-the-loop for dangerous actions.

Advanced Patterns:

• Reflexion: Agent reviews its own output and self-corrects before finalizing
• LATS (Tree Search): Explores multiple action paths and picks the best one
• ReWOO: Plans all tool calls upfront, executes in parallel - faster but less flexible
• Human-in-the-Loop: Agent pauses for human approval before high-stakes actions

Q: What is the difference between ReAct and Plan-and-Execute?

Q: How do you prevent an AI agent from going into infinite loops?

Q: What guardrails would you put on a production AI agent?

Q: How does tool calling work in modern LLMs?