Reading the run result

Runner.run gives back one object, and almost everything you will want to know about a run is on it.

The four you will use

print("final_output:", result.final_output)
print("last_agent:  ", result.last_agent.name)
print("turns used:  ", len(result.raw_responses))

What it cost

The same object carries the token count, one level down, on the context wrapper.

usage = result.context_wrapper.usage

print("requests:", usage.requests)
print("tokens:  ", usage.total_tokens)

requests is how many times a model was called, input_tokens and output_tokens split the cost, and total_tokens adds them up. This is the number that grows quietly as a conversation gets long, which is what lesson 20 is about.

Two requests here and zero tokens, because our stand-in never tokenised anything. Against a real model the requests stay the same and the tokens do not.

Walking the items

Every item has a type, and ItemHelpers pulls the text out of a message without you reaching into the raw shapes.

for item in result.new_items:
    if item.type == "message_output_item":
        print(ItemHelpers.text_message_output(item))

print("final_output:", result.final_output)
print("last_agent:  ", result.last_agent.name)
print("turns used:  ", len(result.raw_responses))
print()
for item in result.new_items:
    print(f"{type(item).__name__:<20} {ItemHelpers.text_message_output(item) if item.type == 'message_output_item' else ''}")

Two turns for a three item run. The first turn produced the tool request, the tool ran without going near the model, and the second turn produced the words. Turns and items are not the same count, and that is worth holding on to when you start reading traces.