Planning and the Agent Loop

An agent works in a loop: it thinks about what to do, takes an action, looks at the result, and repeats until the goal is done.

The Agent Loop

1. Think: "What do I know? What should I do next?"
2. Act: call a tool (search, read, calculate...).
3. Observe: read the tool's result.
4. Repeat or finish with a final answer.

This pattern is often called ReAct (reason + act). In code, the loop is simple. The hard part is the model's decisions inside it.

A Simulated Agent Loop

Here the "model" is a simple Python function so you can run the example without an API. It decides the next action from what it has seen so far:

def fake_model(goal, observations):
    if "price" not in observations:
        return ("call", "get_price", "rice 5kg")
    if "total" not in observations:
        return ("call", "calculator", observations["price"] * 3)
    return ("finish", f"3 bags of rice 5kg cost Rs {observations['total']}")

def run_tool(name, arg):
    if name == "get_price":
        return {"price": 1450}
    if name == "calculator":
        return {"total": arg}

goal = "How much do 3 bags of 5kg rice cost?"
observations = {}
for step in range(1, 6):              # never loop forever
    action = fake_model(goal, observations)
    if action[0] == "finish":
        print("Answer:", action[1])
        break
    _, tool, arg = action
    result = run_tool(tool, arg)
    print(f"Step {step}: {tool}({arg!r}) -> {result}")
    observations.update(result)

Step 1: get_price('rice 5kg') -> {'price': 1450}
Step 2: calculator(4350) -> {'total': 4350}
Answer: 3 bags of rice 5kg cost Rs 4350

A real agent replaces fake_model with a call to a language model, passing the goal, the tool descriptions and the observations so far.

Planning Ahead

For bigger tasks, it helps to ask the model for a plan first, then carry it out step by step:

Goal: Compare 3 laptops under Rs 200,000 for a design student.
First write a numbered plan of at most 5 steps.
Do not start until the plan is written.
After each step, mark it done and update the plan if needed.

A plan might be:

Plans make agents easier to follow and to correct: you can say "skip step 1, here are three models I already like".

Stop Conditions

Agents can get stuck: repeating the same search, or never deciding they are done. Always add limits in your code, not only in the prompt:

• Maximum steps (like range(1, 6) above)
• Maximum time or maximum cost in tokens
• Repeated action check: stop if the same tool is called with the same input twice in a row
• A clear finish rule in the instructions: "When you have the answer, reply with FINAL: and the answer."

history = [("search", "rice price"), ("search", "rice price")]
if len(history) >= 2 and history[-1] == history[-2]:
    print("Stopping: the agent repeated the same action.")

Stopping: the agent repeated the same action.

Reflection

Some agents add a check step: after finishing, the model reviews its own result against the goal ("Did I compare exactly 3 laptops? Are all prices under Rs 200,000?"). This catches many simple mistakes.

Exercise