AI Foundry agent produces unexpected responses

Question

AI Foundry agent produces unexpected responses

Chris 0

I’m running an Azure Web App that uses an AI agent via the Python SDK. However, the agent is currently hardcoded so that its responses should trigger either a blob append or a blob search. Sometimes, nothing happens not even an error.

What is the best way to capture and persist all incoming and outgoing JSON bodies or event traces possibly using Blob Storage or Application Insights for later inspection of message flow anomalies?

Anshika Varshney 12,110 Reputation points Microsoft External Staff Moderator

2025-10-28T18:23:44.99+00:00
Hello Chris,

Thanks for sharing the detailed description that helps a lot! Based on your explanation, it looks like the agent is sometimes not triggering the expected action (append/search blob) and returns no output. This type of issue usually happens when the agent’s internal logic or tool invocation fails silently, or when tracing isn’t enabled to capture what’s going on behind the scenes.

Here are a few steps you can try to identify and fix the issue:

Enable Tracing and Observability

To understand what’s happening inside the agent run, enable tracing in Azure AI Foundry. You can connect your project to Application Insights or Azure Monitor and record all agent activities, model calls, and tool executions.

Go to your project → Tracing tab → connect Application Insights.

In your code, enable content recording:

import os os.environ["AZURE_TRACING_GEN_AI_CONTENT_RECORDING_ENABLED"] = "true"

Reference doc: Trace and observe AI Agents in Azure AI Foundry

Validate tool definitions and permissions

Sometimes the agent’s tools (like Blob append or search) fail silently if there’s a missing permission or configuration issue.

Make sure the agent’s managed identity has correct Blob Storage read/write permissions.

Test each tool independently to confirm they work outside the agent flow.

Review the agent’s system instructions

It’s possible that the agent prompt doesn’t always match a condition to trigger an action. Try explicitly defining fallback behavior like:

"If the request doesn’t match append or search, respond with ‘no action performed’.”

This ensures you always get a visible response rather than silence.

Use tracing to detect ‘no action’ runs

After enabling tracing, you’ll be able to see whether a tool call was made. Look for traces such as tool.call, execute_tool, or agent_planning to confirm what happened during the run.

References:

Azure Storage Blob SDK for Python

Application Insights for Python

Hope this helps clarify the behavior and gives you visibility into what’s happening inside the agent!
Chris 0 Reputation points

2025-10-29T12:05:28.92+00:00

The document you provided doesn’t let me save incoming and outgoing JSON bodies. Is there an easy way to download incoming and outgoing JSON bodies?

1 answer

Your answer

Chris 0 Reputation points

2025-10-29T12:05:28.92+00:00

The document you provided doesn’t let me save incoming and outgoing JSON bodies. Is there an easy way to download incoming and outgoing JSON bodies?

Answer 1

Hello Chris,
You’re right the standard Azure AI Foundry Agent Playground interface doesn't provide a built-in button to directly download the full incoming & outgoing JSON bodies of each run. But yes, there are workarounds and ways to capture them. Here’s how you can do it:

Workaround A: Use Tracing / Observability

You can hook into the agent's tracing/monitoring features so you can see the JSON request/response details (or at least a rich enough log) and then export them manually:

1.In the Azure Portal, go to your Foundry project → Observability or Tracing settings.

2.Connect an Application Insights resource (or Log Analytics workspace) to your agent so that your agent logs are sent to it.

3.Use the SDK or your code to ensure content-recording is enabled for the agent. For example, in Python you may set environment variables or use the SDK to enable recording of messages.

4.In Application Insights / Log Analytics, you can query for request/response spans that include the JSON payloads, then export the logs (CSV/JSON) for offline review.

While this gives you access, it’s not a “one-click download of request/response JSON” from the UI.

Workaround B: Use the SDK/CLI for direct calls

If you call the agent via SDK or REST (instead of only using the Playground), you can capture both the outgoing request (what you send) and the incoming response (what the agent returns) locally. For example:

import json
# build message payload
payload = { "messages": [ {"role":"user","content":"…"} ], /* other params */ }
with open("request.json", "w") as f: f.write(json.dumps(payload, indent=2))

response = client.agents.runs.create_and_process(thread_id=…, agent_id=…)
with open("response.json","w") as f: f.write(json.dumps(response, indent=2))

This way you get both JSON files and can store/analyze them.

Things to note:

The Playground hides many internals for simplicity, so full raw JSON bodies may not always be exposed in UI form.

Logging message content may have privacy/security implications — be careful if your payloads contain sensitive data.

If you only have the UI (no code access) then option A is the only realistic path.

There may be slight delays or sampling in logs, depending on how busy the system is.

So in short: No, there’s no native “Download full JSON” button in the agent Playground. But yes, you can get access to the JSON bodies by:

Enabling tracing/observability and exporting from Application Insights or Log Analytics
Or by using the SDK/REST yourself so you control request & response and write them to files.

I hope this has been helpful! Please let me know if you have another query.

Anshika Varshney 12,110 Reputation points Microsoft External Staff Moderator

2025-10-31T17:03:26.6166667+00:00

Hello Chris,

Following up to see if the below answer was helpful. If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.

Share via

AI Foundry agent produces unexpected responses

1 answer

Your answer