Ease it with AI Agents

Cameras optional. Chat box mandatory.

Today's map

1. 0–10 minWarm-upWho's here, what we're doingchat
2. 10–25 minChatbot vs agentSame question, two very different brainssimulation
3. 25–45 minInside an agent's headBrain, hands, notebook, loop2 simulations
4. 45–65 minAgents at workKlarna, GitHub, Morgan Stanleyquiz
5. 65–70 minBreakWater. Stretch. Unmute nothing.pause
6. 70–100 minLive build: ShopBuddyA support agent from scratch in plain Pythoncode
7. 100–110 minWhen agents go rogueGuardrails, and one very bad day at Replitsimulation
8. 110–120 minYour turn + Q&AWhat you can automate this weekchat

Isn't ChatGPT already an agent? Mostly no. Here's the difference in one picture.

Picture a genius intern. Has read the entire internet. Has no laptop, no notebook, and forgets you every morning. That's a raw LLM. Now let's hire them properly.

The secret sauce isn't a smarter model. It's a loop.

How "using a tool" actually works

The LLM never runs anything. It can only write text. So we show it a menu of tools, and it writes back a polite request. Our code does the real work and hands back the result.

# what the LLM sends back instead of an answer
{"tool_calls": [{"function": {
    "name": "get_order",
    "arguments": "{\"order_id\": \"OD1004\"}"
}}]}
# our Python runs get_order("OD1004"), sends the result back, and the LLM thinks again

Memory: goldfish or elephant?

Without memory, every conversation starts from zero. Production agents keep a small file or database of what matters, then paste the relevant bits back into the prompt next time. That's honestly most of the trick.

memory.json{}

Not demos. Not "coming soon". These are running in production at companies you've heard of.

Also everywhere now: coding agents in your editor, browser agents that click around websites for you, and research agents that read 50 sources while you make chai.

A support agent for Kart, a made-up Indian store. Klarna-style, minus the billion-dollar budget.

1. A chatbot that bluffs

   step1_chatbot.py

   Just an LLM call in a loop. We'll ask where order OD1002 is. It has no idea, but watch how confidently it answers anyway.

   reply = chat(messages)      # no tools, no memory
   print(reply["content"])     # sounds helpful. knows nothing.

   Predict in chat first: will it know where the order is?

2. Give it hands

   tools.py · step2_tools.py

   Four tools: look up an order, check refund policy, issue a refund, raise a ticket for a human. Plus the loop. This is the moment it becomes an agent.

   for step in range(MAX_STEPS):
       reply = chat(messages, tools=TOOLS)          # THINK
       if not reply.get("tool_calls"):
           return reply["content"]                  # done!
       for call in reply["tool_calls"]:
           result = run_tool(call.name, call.args)  # ACT
           messages.append(tool_result(result))     # OBSERVE

   Watch for: it says no to a refund that's outside the 7-day window. Rules, not vibes.

3. Give it a notebook

   memory.py · step3_memory.py

   Memory is a JSON file, and "remembering" is just another tool the agent can choose to call. Then we restart the program and say "hi".

   {"priya@example.com": {
      "facts": ["Prefers WhatsApp updates",
                "Waiting on smartwatch OD1002"] }}

   Watch for: Priya says only "any update?" and it knows exactly what she means.

4. Make it safe enough for real customers

   step4_production.py

   Big refunds pause for a human. Every action goes into an audit log. If Groq goes down, Gemini takes over. A step limit stops runaway loops.

   if refund_amount > 2000:
       ok = input("Supervisor, approve? (y/n): ")
   # the guardrail lives in code, not the prompt.
   # you can't sweet-talk an if-statement.

   Watch for: we'll pull the plug on Groq live and see what happens.

Hands are powerful. Hands can also knock things over.

The guardrails real teams use

Work smarter starting this week. Steal any of these.

Should this be an agent?

Not everything needs one. Ask:

• Does it take several steps, where each step depends on the last?
• Does it need live data or real actions, not just text?
• Does it involve judgment, not just a fixed if-this-then-that rule?

Three yeses: build an agent. Mostly no: a good prompt or a 20-line script will beat it, and cost less.

The whole session in five lines

1. Generative AI answers. Agents act.
2. An agent is a brain, tools, memory and a loop.
3. The LLM only asks for tools. Your code runs them.
4. Memory can be as simple as a JSON file.
5. Guardrails go in code, and humans approve the scary stuff.