Fix as_tool returning blank string on early tool termination

[ihower]

Resolved: #2110

This PR fixes a bug where using Agent.as_tool together with tool_use_behavior causes the lead agent to receive blank string whenever the tool terminates early.

This issue affects not only the case described in #2110 (ToolsToFinalOutputResult), but also stop_on_first_tool and StopAtTools.

Root Cause

as_tool currently returns:

ItemHelpers.text_message_outputs(output.new_items)

This function only extracts text from output.new_items. But when a tool stops early, new_items contains no text messages, so the return value becomes an blank string "".

Solution

Return final_output instead. It already handles early tool-return scenarios correctly.

Behavior Changes

1. text_message_outputs(output.new_items) includes the final assistant message and any intermediate preamble assistant messages.

   • Switching to final_output removes these preambles. Only final assistant message.

   • I think this is fine and reasonable because:

     1. preamble messages are rare unless explicitly prompted
     2. preambles are intended for human readability during tool-calling explanations, and they unnecessarily consume the lead agent’s context window
     3. when preambles exist and the tool terminates early, text_message_outputs(output.new_items) would return the preambles, not actual tool result, which is incorrect anyway.

2. final_output may be structured output instead of a plain string.

   • This aligns with the new tool-call behavior, where tool outputs are no longer restricted to text value and can be structured data (i.e. image or files)

Repro Code

from agents import Agent, function_tool, Runner
from agents.agent import ToolsToFinalOutputResult
import random

@function_tool
def generate_random_number_tool() -> int:
    return random.randint(0, 100)


random_number_agent = Agent(
    name="Sub Agent",
    instructions="You are a helpful assistant that generate_random_number_tool",
    model="gpt-5-mini",
    tools=[generate_random_number_tool],
    tool_use_behavior="stop_on_first_tool"
)

main_agent = Agent(
    name="Main Agent",
    instructions="You are a helpful assistant. Must use the random_number_tool to generate a random number. If tool ouput is empty, just say N/A",
    model="gpt-5-mini",
    tools=[random_number_agent.as_tool(tool_name="random_number_tool", tool_description="A tool that generates a random number.")],
)


result = Runner.run_sync(main_agent, "What is the random number?")
print(result.final_output)

Before fix: You will get "N/A" instead of a random number

[seratch]

Thank you!

[xJkit]

Thank you!