WebSearchTool returns messy, duplicated links and ignores formatting instructions

[Shivam-Bahuguna-Seatrium]

When using WebSearchTool with tool_use_behavior="run_llm_again", the agent’s output does not follow the provided prompt instructions and results in poorly formatted responses.

web_search_tool = WebSearchTool()

 openai_general_search_client = AsyncOpenAI(
    api_key=OPENAI_API_KEY_EUS,
    base_url=f"{OPENAI_ENDPOINT_EUS}openai/v1/"
)
general_search_agent = Agent(
    name="GeneralSearchAgent",
    instructions="""
    Provide:
    - Short summary
    - Bullet points
    - Clean Markdown links
    """,
    model=OpenAIResponsesModel(
        model="<OPENAI_GENERAL_SEARCH_DEPLOYMENT>",
        openai_client=openai_general_search_client
    ),
    tools=[web_search_tool],
    tool_use_behavior="run_llm_again"
)

output be like

[Premier League Table, Standings & Form Guide](https://www.premierleague.com/en/tables)[Live Premier League table: 2025/26 season standings](https://www.skysports.com/football/news/11096/13382434/live-premier-league-table-2025-26-season-standings)[Premier League 2025/2026 - Football League Table and Final Standings ...](https://www.sportinglife.com/football/league-tables/premier-league/1/seasons/2025-2026/107)e the[Premier League Table, Standings & Form Guide](https://www.premierleague.com/en/tables)[Live Premier League table: 2025/26 season standings](https://www.skysports.com/football/news/11096/13382434/live-premier-league-table-2025-26-season-standings)[Premier League 2025/2026 - Football League Table and Final Standings ...](https://www.sportinglife.com/football/league-tables/premier-league/1/seasons/2025-2026/107)tps://www.sportinglife.com/football/league-tables/premier-league/1/seasons/2025-2026/107)

It is happening only when the agent is using the web search tool.. without it, the instructions are being followed.

[seratch]

This looks like the output is trying to follow your instructions. Also, I am not sure what the "OPENAI_GENERAL_SEARCH_DEPLOYMENT" model is, but simple code using a foundation model
such as https://github.com/openai/openai-agents-python/tree/main/examples/financial_research_agent works more intuitively. Please check the difference using the examples in this repo.

[Shivam-Bahuguna-Seatrium]

Sorry My bad! for not providing proper context earlier. I am only using GPT-5 family only.

The difference I noticed is that when I use the streamed runner, I get messy formatting. Without streaming, the output looks fine and properly formatted.

Below is some sample code demonstrating the issue.

try:
   
    # Create WebSearchTool
    web_search_tool = WebSearchTool()
    
    # Create agent with web search capability using async_client
    web_research_agent = Agent(
        name="WebResearchAgent",
        instructions=(
            "You are a research assistant. "
            "Use web search to find accurate, up-to-date information "
            "and summarize it clearly with citations."
            "Ensure proper table formatting with markdown"
        ),
        model=OpenAIResponsesModel(
            model="gpt-5-mini",
            openai_client=async_client 
        ),
        tools=[web_search_tool],
        tool_use_behavior="run_llm_again",
        model_settings=ModelSettings(
            include_usage=True,
        )
    )
    
    
    # Run query 
    query = "give me epl standing table"

    result = await Runner.run(web_research_agent, input=query)
    print("✅ Agent SDK without Streaming Test Results")
    print(result.final_output)
    
    # Use Runner.run_streamed for streaming responses
    result_stream = Runner.run_streamed(web_research_agent, input=query)
    print(f"🔎 Agent SDK Streaming Test Results")

    
    accumulated_output = ""
    
    # Stream events from the agent
    async for event in result_stream.stream_events():
        if event.type == "raw_response_event":
            # Import the type for checking
            from openai.types.responses import ResponseTextDeltaEvent
            
            # Only process text delta events
            if isinstance(event.data, ResponseTextDeltaEvent):
                delta = getattr(event.data, "delta", None)
                if delta and isinstance(delta, str):
                    print(delta, end='', flush=True)
                    accumulated_output += delta
        
    
    elapsed_time = time.time() - start_time
    
    print(f"\n\n{'='*80}")
    print(f"⏱️  Execution Time: {elapsed_time:.2f}s")
    print(f"📊 Output Length: {len(accumulated_output)} characters")

except Exception as e:
    print(f"❌ Error: {str(e)}")
    print(f"   Type: {type(e).__name__}")
    import traceback
    print(f"\n{traceback.format_exc()}")

PRINT:

================================================================================
✅ Agent SDK without Streaming Test Results
================================================================================

Here’s the current English Premier League table (standings) as of 7 January 2026. 

|Pos|Team|Pld|W|D|L|GF|GA|GD|Pts|
|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
|1|Arsenal|20|15|3|2|40|14|+26|48|
|2|Manchester City|20|13|3|4|44|18|+26|42|
|3|Aston Villa|20|13|3|4|33|24|+9|42|
|4|Liverpool|20|10|4|6|32|28|+4|34|
|5|Chelsea|20|8|7|5|33|22|+11|31|
|6|Manchester United|20|8|7|5|34|30|+4|31|
|7|Brentford|20|9|3|8|32|28|+4|30|
|8|Sunderland|20|7|9|4|21|19|+2|30|
|9|Newcastle United|20|8|5|7|28|24|+4|29|
|10|Brighton & Hove Albion|20|7|7|6|30|27|+3|28|
|11|Fulham|20|8|4|8|28|29|-1|28|
|12|Everton|20|8|4|8|22|24|-2|28|
|13|Tottenham Hotspur|20|7|6|7|28|24|+4|27|
|14|Crystal Palace|20|7|6|7|22|23|-1|27|
|15|Bournemouth|20|5|8|7|31|38|-7|23|
|16|Leeds United|20|5|7|8|26|33|-7|22|
|17|Nottingham Forest|21|6|3|12|21|34|-13|21|
|18|West Ham United|21|3|5|13|22|43|-21|14|
|19|Burnley|20|3|3|14|20|39|-19|12|
|20|Wolverhampton Wanderers|20|1|3|16|14|40|-26|6|

Sources: Sky Sports, ESPN, PremierLeague.com. 

Would you like this sorted by form, home/away records, or refreshed for live updates?[Premier League Table, Standings & Form Guide](https://www.premierleague.com/en/tables)[Premier League Table | Sky Sports](https://www.skysports.com/premier-league-table)[2025-26 English Premier League Standings - ESPN](https://www.espn.com/soccer/standings/_/league/eng.1)

================================================================================
🔎 Agent SDK Streaming Test Results
================================================================================

[Premier League Table, Standings & Form Guide](https://www.premierleague.com/en/tables?competition=8&season=2025&round=L_1&matchweek=-1&ha=-1)[2025-26 English Premier League Standings - ESPN](https://www.espn.com/soccer/standings/_/league/eng.1)[Live Premier League table: 2025/26 season standings](https://www.skysports.com/football/news/11096/13382434/live-premier-league-table-2025-26-season-standings)2 |[Premier League Table, Standings & Form Guide](https://www.premierleague.com/en/tables?competition=8&season=2025&round=L_1&matchweek=-1&ha=-1)[2025-26 English Premier League Standings - ESPN](https://www.espn.com/soccer/standings/_/league/eng.1)[Live Premier League table: 2025/26 season standings](https://www.skysports.com/football/news/11096/13382434/live-premier-league-table-2025-26-season-standings)26 season standings](https://www.skysports.com/football/news/11096/13382434/live-premier-league-table-2025-26-season-standings)ndings)son-standings)andings)ndings)ings)-standings)ngs)ndings)season-standings)tandings)ndings)standings)andings)-season-standings)son-standings)ngs)andings)dings)dings)standings)eason-standings)ason-standings)andings)dings)gs)s)ndings)andings)eason-standings)tandings)ngs)s)))s))s)s))s)s)s))s)gs)gs)son-standings)ndings)andings)ings)tandings)andings)ngs)2025-26-season-standings)gs)dings)ings)andings)andings)tandings)26-season-standings)ndings)ings)ings)ings)gs)ngs)gs)ings)ings)eason-standings)gs)on-standings)eason-standings)-table-2025-26-season-standings)son-standings)25-26-season-standings)r-league-table-2025-26-season-standings)2025-26-season-standings)025-26-season-standings)tandings)ngs)standings)s)tandings)-standings)on-standings)ngs)ngs)ndings)ings)tandings)ings)ings)-season-standings)ason-standings)dings)ndings)ndings)ings)andings)andings)son-standings)26-season-standings)eason-standings)eason-standings)son-standings)dings)ings)andings)6-season-standings)n-standings)son-standings)on-standings)ings)ings)ason-standings)eason-standings)season-standings)standings)ngs)eason-standings)season-standings)son-standings)n-standings)ason-standings)tandings)

================================================================================

[Shivam-Bahuguna-Seatrium]

@seratch Hey, can you suggest which event delta can help me in this case? I can see that ResponseTextDeltaEvent is having this type of issue for this case..

[seratch]

I am not sure if this is the right solution you need, but I just asked codex to optimize your code and here is its answer. This is indeed working. You may not need some of the changes, but you can start with this:

---

Hi! I took a swing at a streaming-friendly version that keeps the URLs clean and avoids the double-LLM rewrite that caused the messy citations. Feel free to copy-paste this as-is.

What’s different from your current approach

• Stream as incremental text: consume ResponseTextDeltaEvent chunks and re-render a local buffer on each chunk, rather than expecting a fully formed table to arrive at once (the sync mindset).
• Keep WebSearchTool lightweight with search_context_size="low" so the table stays compact.
• Let the tool’s message be the final table; no post-processing LLM pass.

Tested streaming sample (needs OPENAI_API_KEY)
File: tmp/demo_streaming_table.py

import asyncio
import sys
from agents import Agent, Runner, WebSearchTool
from openai.types.responses import ResponseTextDeltaEvent


def clear_and_render(buffer: str) -> None:
    # Clear screen and show the current markdown table.
    sys.stdout.write("\033[2J\033[H")
    print("Live markdown table (streaming)\n")
    print(buffer)
    sys.stdout.flush()


async def main() -> None:
    agent = Agent(
        name="streaming-table",
        instructions=(
            "Return a 3-row markdown table with columns: Title, Source, Link. "
            "Include a single markdown link in the Link column. Keep it concise."
        ),
        model="gpt-5.1",
        tools=[WebSearchTool(search_context_size="low")],
    )

    result = Runner.run_streamed(
        agent,
        input="List three positive tech news items from today with citations as a markdown table.",
        max_turns=2,
    )

    buffer = ""
    async for event in result.stream_events():
        if event.type == "raw_response_event" and isinstance(event.data, ResponseTextDeltaEvent):
            delta = event.data.delta or ""
            buffer += delta
            if delta.strip():
                clear_and_render(buffer)

    print("\nFinal output:\n")
    print(result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

Runtime behavior

• The table “grows” live as text deltas arrive in the terminal.
• Final markdown table prints at completion with clean, non-duplicated URLs.

[github-actions]

This issue is stale because it has been open for 7 days with no activity.