GPT-5-mini in Azure AI Foundry Agents times out in HIGH reasoning mode but MED has significantly lower quality and slower performance

Hi Elissa,

Thanks for sharing the details. I can see that the timeout is happening when using GPT‑5‑mini with HIGH reasoning effort in Azure AI Foundry Agents.

This behavior is expected in some cases because higher reasoning effort increases processing time, and in multi-step agent workflows the request may exceed the allowed execution time. [ai.azure.com]

Below are some practical steps and examples to help you troubleshoot:

1. Understand why timeout is happening

• HIGH reasoning mode takes more time because the model is doing deeper analysis
• Multi-agent workflows increase total execution time
• If the response takes too long, it can hit timeout limits

Also, synchronous agent or tool calls must complete within a limited time window, otherwise they fail with timeout.

2. Common timeout causes (with examples)

Here are some common real-world causes:

Long reasoning tasks Example: using HIGH reasoning for complex analysis or multi-step logic

Large input payload Example: sending big JSON with many fields or large context

Large output generation Example: generating long responses or detailed reports

Multi-agent orchestration delays Example: step 3 agent waits on step 2 output which itself is slow

Slow external tools or APIs Example: agent calling search API or database that takes time

Waiting for full response instead of partial output Example: backend waits for complete response before returning

High latency or region load Example: same request works sometimes but times out during peak load

Timeout can also happen due to network latency or service overload in backend calls.

3. Enable streaming response (very important)

Right now you are waiting for full completion. This increases timeout risk.

What streaming does:

• Returns output in chunks as soon as it is generated
• User sees partial response immediately
• Reduces perceived timeout issues

How to enable (general approach):

• If using SDK
  • Enable streaming option in response call
    • Example concept stream = true
    • If using API
      • Use streaming-supported endpoint or parameter
      • If using UI or Playground
        • Turn on streaming or real-time output option

Streaming is recommended for long responses because it improves responsiveness while the model is still processing. (Streaming responses are supported for real-time interaction in Foundry agents.) [drafts.cod...thme.cloud]

4. Reduce reasoning effort strategically

Instead of using HIGH everywhere:

• Use MED or LOW for most steps
• Use HIGH only where deep reasoning is needed
• Combine models if needed (fast + reasoning mix)

This helps balance quality and performance.

5. Break large tasks into smaller steps

Avoid doing everything in one request.

Example:

• Step 1: preprocess input
• Step 2: generate intermediate output
• Step 3: final reasoning

This keeps each step within execution limits.

6. Optimize token usage

Latency depends on tokens processed.

• Reduce unnecessary input data
• Avoid repeating context
• Limit response length

More tokens mean more processing time.

7. Check slow steps in workflow

If your agent uses tools or sub-agents:

• Identify which step is slow
• Check logs or traces
• Optimize that specific part

Often timeout is caused by one slow dependency.

8. Check region and performance factors

• Try another region if possible
• Monitor if issue is consistent or intermittent
• Check if happens in both portal and backend

In short: This timeout happens mainly due to:

• high reasoning effort
• long-running workflows
• large input or output
• slow tool execution

Best approach:

• use streaming responses
• reduce reasoning where possible
• break workflow into smaller steps

I Hope this helps. Do let me know if you have any further queries.

Thankyou!

Hi Elissa,

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thankyou!

Hi Elissa,

Just checking back to see if you’re still facing the same issue. If the problem persists, please share a few more details and we’ll be happy to help you further.

Thankyou!

Hello @Anshika Varshney , hope you are doing well. These are the actions we have been able to take based on your previous feedback:

1. Understand why timeout is happening

This has been difficult for us since timeouts are happening directly in foundry portal using the agent as a single agent. So we don't believe it is related to the multi-step orchestration.

2. Common timeout causes (with examples)

Our input is generally a JSON file that is not extremely long and as mentioned before, it is only causing timeouts with the reasoning effort set to HIGH. The output it should generate is a JSON that is much shorter than the input. In our case it is sequential, but the agents processing doesn't begin until the previous one is finished. We are not calling any external tools and have isolated the issue to the agent that has the gpt-5-mini.

3. Enable streaming response (very important)

We looked for this option within the SDK and were not able to enable it. Could you provide further guidance?

4. Reduce reasoning effort strategically

Our agent is the last step of the process, and as mentioned before, using MED or LOW reasoning effort has reduced our quality significantly.

5. Break large tasks into smaller steps

This is already happening in our workstream. The agent with gpt-5-mini is the last one in the process in charge.

6. Optimize token usage

There is only one call to this agent, so there is no repeated context given anywhere, it's the first and only call we make per process.

7. Check slow steps in workflow

There are no external tools used, the agent is timing out without dependencies.

8. Check region and performance factors

Changing regions is not possible for us because we have many other agents hosted in this same region and subscription.

If you have any other recommendations on how we can troubleshoot this issue it would be highly appreciated.

We also wanted to know, is there any way we can keep ourselves informed about recent changes in foundry models or other updates on this tool? We would like to be able to mitigate the risks of these changes within models affecting our workflow as much as possible.

Thanks again for your help.

Thank you for the detailed update, this really helps to narrow things down. I understand this is happening even with a single agent in the portal and only when using high reasoning, which makes this a bit different from typical timeout scenarios.

Based on what you shared, here are a few additional things you can try:

1. About timeouts with high reasoning

With high reasoning, the model takes more time internally to process and generate the answer.

Even if your input and output are not very large, the internal reasoning steps can still be heavy, which can lead to timeouts in the portal itself.

Since you already confirmed this happens only with high reasoning, this strongly suggests it is computing time related rather than input size or workflow issues.

2. Try reducing max output tokens

Even if your expected output is small, the model may still reserve space for larger responses.

You can try:

• explicitly setting a lower max tokens value
• for example around 200 to 500 depending on your output

This helps reduce overall processing time and can avoid timeouts.

3. Simplify the instruction slightly

Even small changes in instructions can impact reasoning time.

You can try:

• making the instruction more direct
• removing unnecessary explanation or context
• guiding the model to produce structured output faster

This helps the model spend less time in reasoning loops.

4. Streaming support clarification

Right now, streaming is not always exposed directly in the Foundry portal experience.

Streaming is typically used in API or SDK calls, not in the portal UI itself.

So, if you are testing inside the portal, this is expected that you do not see a streaming option there.

5. Try testing via API

If possible, try calling the same deployment using API instead of only the portal.

This helps confirm:

• whether the timeout is UI related
• or truly coming from the model execution

Sometimes portal limits are stricter compared to API calls.

6. Check token usage in practice

Even if you believe input is small, it is helpful to verify:

• full JSON size in tokens
• hidden system instructions added by the agent
• any formatting or schema instructions

These can silently increase total tokens and processing time.

7. Model behavior note

For GPT-5 mini and similar models:

• high reasoning mode increases latency significantly
• this is expected behavior for complex reasoning tasks
• some scenarios may hit time limits even with moderate input

8. About staying updated on changes

This is a great question.

You can track updates here:

• Azure AI Foundry documentation
• https://ai.azure.com/catalog/models

These pages are updated regularly with new features and model changes.

I hope this helps you get back on track! If you're still facing issues, could you share more details?

Thankyou!