Chain-of-thought (CoT) prompting asks a language model to work through intermediate steps before giving its answer. Instead of requesting only a result, give it a short reasoning procedure: identify the relevant facts, work through them in order, then check the result. This is useful when a plausible first answer may overlook part of a multistep problem. section 5.1.1 section 5.1
Count the time spent in the past
Suppose a traveler spends six days in one past year, jumps forward 50 years and spends another 20 days there, then jumps back to witness an event’s end. A quick answer might count only the first visit. A CoT prompt instead asks the model to map each jump, decide which stays count as being in the past, add their durations, and verify what it included. Both stays count, giving 6 + 20 = 26 days. section 5.1.1
Listing 5.2 01_demonstrating_CoT_reasoning.md (continued)
#With reasoning - append to prompt CoT Reasoning Strategy Step 1: Define present day as Year 0, past as negative years. Step 2: Map traveler's journey: where he goes and how long stays. Step 3: Identify which time periods count as "in the past." from the travelers perspective. Step 4: Track days spent during each visit to past time periods. Step 5: Add up all days spent in past before event ends. Step 6: Verify calculation includes all past time, excludes present/future time.
A reusable six-step prompt for the time-travel problem, from defining the timeline through checking which days belong in the total. section 5.1.1
For a small agent, the instruction can be as simple as “Work out the solution step by step, then give the final answer.” In a year-jump example, the resulting steps subtract ten years from 2025, then add five, reaching 2020. The steps make the calculation inspectable; they do not guarantee it is correct. section 5.2.1 section 5.6
Listing 5.5 02_CoT_agent.py
cot_agent = Agent(
name="TimeTravelerCoT",
instructions=( #1
"You are a time travel problem solver. "
"Work out the solution step by step, then give the final answer."
),
)
question = ( #2
"Starting in 2025, you travel 10 years to the past, then 5 years to the future. "
"What year do you end up in?"
)
result = asyncio.run( #3
Runner.run( #4
cot_agent,
input=question))
print(result.final_output)
OUTPUT
Let's solve this step by step:
1. **Starting Year**: 2025
2. **Travel 10 Years to the Past**:
- 2025 - 10 = 2015
3. **Travel 5 Years to the Future**:
- 2015 + 5 = 2020
A minimal OpenAI Agents SDK example showing where the CoT instruction goes and what the agent returns. section 5.2.1
Choose the amount of reasoning structure to match the task:
- Use CoT for multistep questions with intermediate results. It can discourage shallow shortcuts, but it does not add reasoning ability to an underlying model. section 5.1 section 5.3
- Skip explicit CoT for a reliable single-step lookup or fixed-format response; the extra steps cost tokens and latency. section 5.3 section 5.6
- When the agent must call tools and react to their results, extend the approach with ReAct. Validate reasoning output before relying on it in later steps. section 5.6
