Skip to main content

​
What is an AI Agent?

Let’s start with the basics. You’ve been using LLMs for single-shot tasks:
Pseudocode
// Simple LLM call - one question, one answer
const response = await llm.chat({
    model: "claude-sonnet-4-20250514",
    messages: [{ role: "user", content: "What's the capital of France?" }]
});
// → "Paris"
This works great for simple questions. But what if you need to:
  • Look up information the LLM doesn’t know (customer data, inventory levels)
  • Take actions in external systems (create tickets, send emails, update databases)
  • Remember context across multiple turns (maintain conversation state)
  • Make decisions based on intermediate results (if X then Y)
That’s what agents do. Definition: An AI agent is an LLM system that can:
  1. Perceive - Understand the current situation (user request, context, state)
  2. Think - Reason about what to do (using the LLM)
  3. Act - Execute actions using tools (function calls, API calls)
  4. Remember - Maintain state and context
  5. Repeat - Continue until task is complete

​
The Basic Agent Loop

+-------------------------------------+
│ 1. User Query                       │
+-------------------------------------+
                 |
                 v
+-------------------------------------+
│ 2. Agent Thinks                     │
│   "I need customer data to answer"  │
+-------------------------------------+
                 |
                 v
+-------------------------------------+
│ 3. Agent Acts                       │
│    Calls: get_customer(id="123")    │
+-------------------------------------+
                 |
                 v
+-------------------------------------+
│ 4. Tool Returns Data                │
│    {name: "Alice", orders: 5}       │
+-------------------------------------+
                 |
                 v
+-------------------------------------+
│ 5. Agent Thinks Again               │
│    "Now I can answer"               │
+-------------------------------------+
                 |
                 v
+-------------------------------------+
│ 6. Agent Responds                   │
│    "Alice has placed 5 orders"      │
+-------------------------------------+

​
Your First Agent: Simple Tool Use

Let’s build the simplest possible agent - one that can look up weather information. What just happened:
  1. User asks about weather
  2. Agent sees it has a get_weather tool available
  3. Agent decides to use that tool
  4. We execute the tool and give result back to agent
  5. Agent uses the result to formulate final answer
Key Insight: The agent made a decision. We didn’t tell it “call get_weather.” It reasoned that to answer the question, it needed weather data, and chose the appropriate tool.

​
When to Use Agents vs. Simple LLM Calls

Use a simple LLM call when:
  • ✅ Single question with immediate answer
  • ✅ No external data needed
  • ✅ No actions to take
  • ✅ No context to maintain
Use an agent when:
  • ✅ Need to look up external data (databases, APIs)
  • ✅ Need to take actions (create records, send messages)
  • ✅ Multi-turn conversation with context
  • ✅ Multi-step reasoning required
  • ✅ Decision making based on intermediate results
Real-World Examples:
TaskApproachWhy
”Summarize this document”Simple LLMOne-shot, no tools needed
”What’s my account balance?”AgentNeeds to query database
”Cancel my subscription”AgentNeeds to execute action
”Help me debug this code”Agent (if using tools)May need to run code, search docs
”Translate this text”Simple LLMOne-shot, no tools needed

​
Agent Limitations and Considerations

Agents are powerful but come with challenges: 1. Unpredictability
  • Same query might take different paths
  • Agent might call tools in unexpected order
  • Need to handle variability
2. Cost
  • Multiple LLM calls per query (thinking → acting → thinking)
  • Tools add latency and potential API costs
  • Need to optimize for efficiency
3. Error Handling
  • Tools can fail
  • Agent might choose wrong tool
  • Need robust error recovery
4. Testing Complexity
  • Can’t just test input → output
  • Must test tool selection, execution paths
  • Need to mock tools for testing
The rest of this module teaches you how to handle these challenges in production.

​
Check Your Understanding

  1. Concept: What’s the difference between a simple LLM call and an agent?
  2. Application: User asks “Send an email to my team with today’s sales numbers.” Do you need an agent?