From Chatbots to Agents

1# The classic chatbot loop
2while True:
3    user_message = input("You: ")
4
5    response = llm.generate(
6        system="You are a helpful assistant.",
7        messages=[{"role": "user", "content": user_message}]
8    )
9
10    print(f"Bot: {response}")
11
12# That's it. Receive text, produce text. Nothing more.

1User: What's the current price of Bitcoin?
2
3Chatbot: I apologize, but I don't have access to real-time
4data. My knowledge was last updated in April 2024. For
5current Bitcoin prices, please check a financial website.
6
7# The chatbot KNOWS it should check a website
8# But it CANNOT actually do it

1User: Write a Python function to calculate Fibonacci numbers.
2
3Chatbot: Here's the function:
4
5def fibonacci(n):
6    if n <= 1:
7        return n
8    return fibonacci(n-1) + fibonacci(n-2)
9
10# This is correct, but the chatbot cannot:
11# - Run the code to verify it works
12# - Test edge cases
13# - Benchmark performance
14# - Fix bugs if any existed

1# The agent loop - completely different from chatbots
2def agent_loop(goal: str):
3    state = initialize_state(goal)
4
5    while not goal_achieved(state):
6        # 1. PERCEIVE: Gather context
7        context = gather_context(state)
8
9        # 2. REASON: Decide what to do next
10        action = llm.decide_action(context, goal)
11
12        # 3. ACT: Execute the action
13        result = execute_tool(action)
14
15        # 4. OBSERVE: Update state with results
16        state = update_state(state, result)
17
18        # 5. REFLECT: Learn from the outcome
19        if needs_replanning(state, result):
20            state = replan(state, goal)
21
22    return state.final_result
23
24# Example: "Research competitors and create a report"
25result = agent_loop("Research top 3 competitors and create a comparison report")
26# The agent will:
27# - Search the web for competitor information
28# - Read and extract relevant data
29# - Compare features and pricing
30# - Generate and save a formatted report
31# - All autonomously!

1# CHATBOT APPROACH
2
3User: There's a bug in my sort function. It's not working.
4
5Chatbot: I'd be happy to help! Could you share the code
6so I can take a look?
7
8User: [pastes code]
9
10Chatbot: I see the issue. On line 15, you're using <=
11instead of <. This causes an off-by-one error. Here's
12the corrected version:
13
14[corrected code]
15
16Try this and let me know if it works!
17
18# Problems:
19# - User must identify there IS a bug
20# - User must find and share the code
21# - Chatbot cannot verify the fix actually works
22# - User must apply the fix manually
23# - If wrong, entire process repeats

1# AGENT APPROACH
2
3User: Find and fix bugs in my sort function.
4
5Agent: I'll help you debug the sort function. Let me:
61. First, find the sort function in your codebase
7   [Searches files...]
8   Found: src/utils/sort.ts
9
102. Read and analyze the code
11   [Reading file...]
12   Found potential issue on line 15: off-by-one error
13
143. Run existing tests to confirm the bug
15   [Executing tests...]
16   FAILED: test_sort_edge_cases - confirms the issue
17
184. Apply the fix
19   [Editing file...]
20   Changed line 15: <= to <
21
225. Verify the fix
23   [Executing tests...]
24   PASSED: All 12 tests passing
25
26Done! Fixed an off-by-one error on line 15 of src/utils/sort.ts.
27The bug caused incorrect sorting when adjacent elements were equal.
28All tests now pass.
29
30# Benefits:
31# - Agent found the code automatically
32# - Agent verified the bug existed
33# - Agent fixed AND verified the fix
34# - User got a working solution, not just advice

1# Modern LLMs have native tool calling
2response = client.messages.create(
3    model="claude-sonnet-4-20250514",
4    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
5    tools=[
6        {
7            "name": "get_weather",
8            "description": "Get current weather for a city",
9            "input_schema": {
10                "type": "object",
11                "properties": {
12                    "city": {"type": "string"}
13                }
14            }
15        }
16    ]
17)
18
19# The model now returns structured tool calls:
20# {
21#     "type": "tool_use",
22#     "name": "get_weather",
23#     "input": {"city": "Tokyo"}
24# }

1# Claude Code in action
2$ claude "Add input validation to the user registration form"
3
4Claude Code will:
51. Search your codebase for the registration form
62. Analyze the existing code structure
73. Implement appropriate validation
84. Write tests for the validation
95. Run tests to verify everything works
10
11# All without you specifying file paths or providing context