Inspecting and Understanding Agent Result Properties

Introduction And Lesson Overview

Welcome back! In the previous lesson, you learned how to define and run your first OpenAI agent using the Agents SDK . You also saw how to extract the agent’s final output from the result object. In this lesson, we will take a closer look at the result object that is returned after running an agent. Understanding the structure and properties of this object is key to building more advanced applications, debugging your agent’s behavior, and making the most of the SDK’s features. By the end of this lesson, you will know how to inspect and interpret the different attributes of the result object, including the final output, the original input, the last agent that ran, new items generated during the run, and the raw responses from the language model. You will also see how these properties can help you understand what happened during the agent’s run and how to use this information in your own projects.

The Structure Of The Result Object

When you run an agent synchronously (using Runner.run_sync) or asynchronously (using Runner.run), you receive a RunResult object. If you run an agent in streamed mode (using Runner.run_streamed), you receive a RunResultStreaming object instead. Both types of result objects provide detailed information about the agent’s execution. Some of the most relevant properties include: final_output : The final output produced by the last agent that ran. Its type can vary depending on the agent’s configuration—it may be a string or a more complex object if the agent specifies an output_type . input : The original input or prompt provided to the agent at the start of the run. last_agent : The agent instance that produced the final output. This is especially useful in workflows involving multiple agents or handoffs. new_items : A list of items generated during the run, such as messages, tool calls, or handoffs. These items provide a step-by-step record of the agent’s reasoning and actions. raw_responses : The raw outputs from the language model for each step in the agent loop, including generated text, tool calls, and finish reasons. By inspecting these properties, you gain a comprehensive view of the agent’s execution, including the input, output and intermediate steps. This structure is designed to support both simple and advanced use cases, from basic logging to complex multi-agent workflows.

Join the 1M+ learners on CodeSignal

Be a part of our community of 1M+ users who develop and demonstrate their skills on CodeSignal

Inspecting The Final Output Property

As a reminder from the previous lesson, the most important property in the result object is usually final_output. This is the agent’s final answer after completing its reasoning, tool use, or any handoffs to other agents. You can access it directly with result.final_output.

For example, after running our Travel Genie agent:

Python

# Print the agent's final output
print("Final Output:", result.final_output)

# Print the agent's final output
print("Final Output:", result.final_output)

You might see output like this:

text

Final Output: Absolutely! For adventure seekers, **New Zealand** is a top pick. This country is a playground for thrill-seekers, offering everything from **bungee jumping in Queenstown** (the "adventure capital of the world"), **skydiving over Lake Taupo**, **white-water rafting in Rotorua**, to **hiking the Tongariro Alpine Crossing**.

**Why it's great for adventure lovers:**
- **Variety:** Mountains, forests, glaciers, beaches—all packed into a compact country.
- **World-class operators:** Safety is prioritized, and there are plenty of reputable guides for every activity.
- **Unique experiences:** Try the original commercial bungee jump, conquer icy crevasses on the Franz Josef Glacier, or go black-water rafting in the Waitomo Glowworm Caves.

**Travel Tips:**
- **Best time to visit:** December to March (summer), but winter brings amazing skiing and snowboarding.
- **Getting around:** Consider renting a camper van for maximum flexibility.
- **Book ahead:** Popular activities like glacier hikes and skydives can sell out, so reserve early.

Other awesome contenders:
- **Costa Rica** for jungle zip lines, volcano hikes, and surfing.
- **South Africa** for shark cage diving, safaris, and mountain biking.
- **Iceland** for glacier trekking, ice cave exploration, and volcanic landscapes.

Let me know what kind of adventure excites you most, and I can tailor recommendations just for you!

Final Output: Absolutely! For adventure seekers, **New Zealand** is a top pick. This country is a playground for thrill-seekers, offering everything from **bungee jumping in Queenstown** (the "adventure capital of the world"), **skydiving over Lake Taupo**, **white-water rafting in Rotorua**, to **hiking the Tongariro Alpine Crossing**.

**Why it's great for adventure lovers:**
- **Variety:** Mountains, forests, glaciers, beaches—all packed into a compact country.
- **World-class operators:** Safety is prioritized, and there are plenty of reputable guides for every activity.
- **Unique experiences:** Try the original commercial bungee jump, conquer icy crevasses on the Franz Josef Glacier, or go black-water rafting in the Waitomo Glowworm Caves.

**Travel Tips:**
- **Best time to visit:** December to March (summer), but winter brings amazing skiing and snowboarding.
- **Getting around:** Consider renting a camper van for maximum flexibility.
- **Book ahead:** Popular activities like glacier hikes and skydives can sell out, so reserve early.

Other awesome contenders:
- **Costa Rica** for jungle zip lines, volcano hikes, and surfing.
- **South Africa** for shark cage diving, safaris, and mountain biking.
- **Iceland** for glacier trekking, ice cave exploration, and volcanic landscapes.

Let me know what kind of adventure excites you most, and I can tailor recommendations just for you!

The final_output property contains the complete, formatted response from the agent. In more complex scenarios, the result object may also include a finish reason, which explains why the agent stopped (for example, because it reached a final answer, hit a token limit, or encountered a content filter). This information can be useful for debugging or for handling special cases in your application.

Analyzing Raw Responses

The raw_responses property is another valuable part of the result object. It contains the raw outputs from the language model for each step in the agent loop. These responses include detailed information such as the generated text, the role (assistant or tool), the status, and sometimes the finish reason (for example, "stop" or "length").

Inspecting raw_responses is especially useful for debugging or for understanding exactly how the language model responded at each step. You can also see if the agent made any tool calls, what the tool outputs were, and how the agent processed those results.

Here is how you might print the raw responses:

Python

# Print raw responses from the LLM
print("Raw Responses:")
for response in result.raw_responses:
    print(f"  - {response}")

# Print raw responses from the LLM
print("Raw Responses:")
for response in result.raw_responses:
    print(f"  - {response}")

A sample output could look like this:

text

Raw Responses:
  - ModelResponse(output=[ResponseOutputMessage(id='msg_682306258fac8198afb05543a94d7a14063efd0493fd7c51', content=[ResponseOutputText(annotations=[], text='Absolutely! For adventure seekers, **New Zealand** is a top pick. ...', type='output_text')], role='assistant', status='completed', type='message')], usage=Usage(requests=1, input_tokens=40, output_tokens=301, total_tokens=341), response_id='resp_68230624da2881988611e4c883ce653a063efd0493fd7c51')

Raw Responses:
  - ModelResponse(output=[ResponseOutputMessage(id='msg_682306258fac8198afb05543a94d7a14063efd0493fd7c51', content=[ResponseOutputText(annotations=[], text='Absolutely! For adventure seekers, **New Zealand** is a top pick. ...', type='output_text')], role='assistant', status='completed', type='message')], usage=Usage(requests=1, input_tokens=40, output_tokens=301, total_tokens=341), response_id='resp_68230624da2881988611e4c883ce653a063efd0493fd7c51')

By reviewing the raw_responses, you can see the full details of each step, including the text generated, the number of tokens used, and the finish reason. This level of detail is helpful for troubleshooting, optimizing your agent’s instructions, or simply understanding how the agent thinks.

Introduction And Lesson Overview

The Structure Of The Result Object

Join the 1M+ learners on CodeSignal

Creating An Agent For Our Examples

Inspecting The Final Output Property

Inspecting the Original Input Property

Inspecting the Last Agent Property

Inspecting the New Items Property

Analyzing Raw Responses

Lesson Summary And Next Steps