Module 08: AI Agents | Generative AI & LLM

Duration: 4 hours | Difficulty: Intermediate–Advanced | Prerequisites: Modules 01–04

Learning Objectives

By the end of this module, you will be able to:

Explain what AI agents are and how they differ from simple LLM chains
Implement the ReAct (Reason + Act) pattern from scratch
Build agents with tool use and function calling
Design agent loops with proper planning, execution, and reflection
Handle errors and edge cases in agent systems

1. What Are AI Agents?

From Chains to Agents

A chain is a fixed recipe. An agent picks steps at runtime: read state → think → maybe call a tool → repeat. Think improvising cook vs meal kit with numbered bags.

Chain:     Step 1 → Step 2 → Step 3 → Done  (fixed path)

Agent:     Observe → Think → Act → Observe → Think → Act → ... → Done
           (dynamic path, decided at runtime by the LLM)

The Agent Loop

┌─────────────────────────────────────────┐
│              AGENT LOOP                  │
│                                          │
│  1. OBSERVE: What's the current state?   │
│       ↓                                  │
│  2. THINK: What should I do next?        │
│       ↓                                  │
│  3. ACT: Execute a tool or respond       │
│       ↓                                  │
│  4. Check: Am I done?                    │
│       ↓                                  │
│     No → Go to step 1                    │
│     Yes → Return final answer            │
└─────────────────────────────────────────┘

Chain	Agent
Predictable	Flexible
Easier to test	Needs budgets + tracing
Great for ETL-ish flows	Great for open-ended tasks

Fun Fact: Most “agents” in prod are small finite-state machines with an LLM inside—not sci-fi autonomy.

ReAct interleaves natural-language thought with actions (tool calls) and observations (tool results). Older tutorials used Thought: / Action: / Observation: text parsing; modern APIs prefer structured tool calls—but the idea stays: reason, act, observe, repeat.

Piece	Role
Thought	Plan visible (helps debugging)
Action	Tool name + args
Observation	Ground truth from your code

Try This! Log every observation in JSON—your future self is debugging at 2 a.m.

Key Takeaway

ReAct is really explicit scaffolding so the model doesn’t “silent tool call” in free text.

3. Tool Use with Function Calling

Providers expose JSON schemas; the model returns tool_calls you execute in Python and feed back as role: "tool" messages.

Key Example: Minimal loop: assistant requests tools until it returns plain text.

python

[object Object], json
,[object Object], openai ,[object Object], OpenAI

client = OpenAI(api_key=,[object Object],)


,[object Object], ,[object Object],(,[object Object],) -> ,[object Object],:
    data = {
        ,[object Object],: {,[object Object],: ,[object Object],, ,[object Object],: ,[object Object],},
        ,[object Object],: {,[object Object],: ,[object Object],, ,[object Object],: ,[object Object],},
    }
    ,[object Object], data.get(city, {,[object Object],: ,[object Object],, ,[object Object],: ,[object Object],})


TOOLS = [
    {
        ,[object Object],: ,[object Object],,
        ,[object Object],: {
            ,[object Object],: ,[object Object],,
            ,[object Object],: ,[object Object],,
            ,[object Object],: {
                ,[object Object],: ,[object Object],,
                ,[object Object],: {,[object Object],: {,[object Object],: ,[object Object],}},
                ,[object Object],: [,[object Object],],
            },
        },
    }
]

messages = [
    {,[object Object],: ,[object Object],, ,[object Object],: ,[object Object],},
    {,[object Object],: ,[object Object],, ,[object Object],: ,[object Object],},
]

,[object Object], ,[object Object],:
    resp = client.chat.completions.create(
        model=,[object Object],, messages=messages, tools=TOOLS, tool_choice=,[object Object],
    )
    msg = resp.choices[,[object Object],].message
    messages.append(msg)
    ,[object Object], ,[object Object], msg.tool_calls:
        ,[object Object],(msg.content)
        ,[object Object],
    ,[object Object], call ,[object Object], msg.tool_calls:
        args = json.loads(call.function.arguments)
        result = get_weather(**args) ,[object Object], call.function.name == ,[object Object], ,[object Object], {,[object Object],: ,[object Object],}
        messages.append(
            {,[object Object],: ,[object Object],, ,[object Object],: call.,[object Object],, ,[object Object],: json.dumps(result)}
        )

4. Agent Memory

Agents need short-term (chat history), working (current plan, variables), and sometimes long-term (notes DB, vector store).

Memory kind	Implementation sketch
Short-term	Trim/summarize last N messages
Working	Dict in your orchestrator; inject as system preamble
Long-term	Embeddings + DB; retrieve like RAG

Concept: Chat transcripts are not a database—persist what matters explicitly.

5. Planning Strategies

Plan-and-Execute

Phase 1: ask the LLM for a numbered plan with tool hints. Phase 2: execute steps sequentially (or parallelize when safe). Phase 3: synthesize final answer. Separates strategy from tactics.

When it helps	When it’s overkill
Multi-tool workflows	Single API lookup

6. Error Handling in Agent Systems

Failure	Prod response
Tool timeout	Retry with backoff; surface friendly error
Bad JSON args	Re-prompt model with validation message
Infinite loop	`max_iterations` + circuit breaker
Model refuses	Fallback model or degrade to search-only

Plain-English robust pattern: wrap each tool in try/except, append structured errors to messages, let the model recover once or twice, then bail with a safe template response.

Key Takeaway