Understanding AI Agents Through Practical Examples
Distinguish chat, a fixed workflow, and an agent through a fictional stock lookup with observations and stopping rules.
Overview
For learners who want to complete this practical exercise with inspectable sample data and explicit checks.
- Map the goal, decision, tool, and observation loop.
- Distinguish a proposed action from application execution.
- Set turn budgets and permission boundaries.
You will build: A trace containing a tool, observation, answer, and explicit stopping condition.
Prerequisites
- A text editor; optionally Python 3 for the simulation.
The conceptual exercise and simulation run offline without an account or API. A fixed-rule simulation must not be presented as an LLM agent.
Step-by-step tutorial
Step 1. Compare three ways to solve one task
Task: report DEMO-PEN stock in a training inventory. Chat without tools can only use conversation data. A fixed workflow always executes its prescribed sequence. An agent chooses actions based on a goal and observations and may ask for clarification or stop. Merely adding an LLM does not turn every workflow into an agent.
Step 2. Define permissions and data
Goal: answer the training stock question
Allowed tool: lookup_stock(sku), read-only
Fixture: DEMO-PEN = 7; DEMO-BOOK = 0
Unknown SKU: missing, never zero by assumption
Budget: at most 3 decisions
Forbidden: purchases, database writes, shell commandsDistinguish zero known stock from missing data. For changing facts, a tool observation is preferable to model memory, but the tool still needs permission and source checks.
Step 3. Run an explicitly labeled simulation
stock = {"DEMO-PEN": 7, "DEMO-BOOK": 0}
question_sku = "DEMO-PEN"
observation = None
for turn in range(3):
if observation is None:
observation = {"found": question_sku in stock,
"units": stock.get(question_sku)}
print("TOOL", observation)
else:
answer = str(observation["units"]) if observation["found"] else "unknown"
print("FINAL fictional stock:", answer)
break
else:
raise RuntimeError("turn limit")This is a rule-based controller without model inference or API calls. Use it to read a trace and test stopping before the next tutorial’s actual model-driven Python agent. Change the SKU to DEMO-BOOK and then MISSING to compare zero with unknown.
Step 4. Read the trace and identify human decisions
Record each input, proposed tool, result, and stop condition. A change from “check stock” to “purchase” requires a new permission; do not silently add a writing tool. If the SKU is missing, request it rather than guessing a similar product.
Review this trace. Identify tool-supplied data, controller decisions, and cases that must stop or ask for clarification. Add no purchasing or data-writing actions.Verification
- Identify the simulation and do not claim it uses an LLM.
- DEMO-BOOK returns 0, MISSING returns unknown, and execution is bounded.
Troubleshooting
Every chatbot is called an agent
Cause: No action/observation loop has been identified.
Fix: Map actual steps and identify where a model can choose an action.
Missing data becomes zero
Cause: The found state is conflated with units.
Fix: Keep found=false and return unknown instead of inferring stock.
Exercises & extensions
- Write a trace for two SKUs and define its maximum tool-call count.
- Continue to the Python agent to replace fixed decisions with the Responses API.
The simulation measures no model autonomy, accuracy, or reliability. Every inventory value is fictional.