Manual ReAct Trace: Walking Through One Query Step by Step

Before automating anything, it's worth manually driving the ReAct loop one step at a time. This is exactly what this course's walkthrough does first: by running the logic cell-by-cell (or, outside a notebook, one Python statement at a time), you get full visibility and control over the agent's thinking process, which makes it far easier to debug and validate its behavior at each individual step before trusting an automated controller to drive it unattended.

The manual process for a question like 'What is India's population divided by 2?' looks like this: you instantiate the Agent with the ReAct system prompt, then call it with the user's question. The model responds with a Thought (recognizing it needs to look up India's population) followed by PAUSE — and then stops, waiting for you.

At this point, you call the agent again, but with a blank message (an empty string) to let it continue reasoning from where it left off. It responds with an Action line — 'Action: lookup_population: India' — followed by another PAUSE. Now it's genuinely waiting on you: you, the human controller in this manual version, actually run the lookup_population('India') tool function yourself, get back a population number, and call the agent again — this time passing in 'Observation: 1417492000' as the message, exactly matching the format the system prompt told the model to expect.

The agent then continues from that observation, does the division, and outputs the final 'Answer: 708746000' line, at which point the loop is done. Running through this by hand — reading the printed output at every single step — is what makes the mechanics of the ReAct loop concrete before you hand control over to an automated agent_loop() function, which is exactly what happens next.

def math(expr: str) -> float:
    return eval(expr)  # fine for a controlled demo; never eval untrusted input in prod

def lookup_population(country: str) -> int:
    return {"India": 1_417_492_000, "USA": 341_814_000}.get(country, 0)

# --- manual, step-by-step ReAct session ---
agent = Agent(system=REACT_SYSTEM_PROMPT)

print(agent("What is India's population divided by 2?"))
# -> "Thought: I need to look up India's population first.\nPAUSE"

print(agent(""))  # blank input lets the agent continue to the next stage
# -> "Action: lookup_population: India\nPAUSE"

# The human (controller) actually runs the tool here:
result = lookup_population("India")

print(agent(f"Observation: {result}"))
# -> "Thought: Now I can compute half of that.\nAction: math: 1417492000 / 2\nPAUSE"

result = math("1417492000 / 2")
print(agent(f"Observation: {result}"))
# -> "Answer: 708746000.0"