[BUG] AgentCoreBrowser empty message

[jmorgan-restrata]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.20

Tools Package Version

0.2.3

Tools used

AgentCoreBrowser

Python Version

3.11.13

Operating System

Linux

Installation Method

pip

Steps to Reproduce

N/A

Expected Behavior

N/A

Actual Behavior

N/A

Additional Context

Our agent performed some research using AgentCoreBrowser. The agent then responded with the research. On attempting to continue the interaction we received the following error message:

Error: An error occurred (ValidationException) when calling the ConverseStream operation: The content field in the Message object at messages.23 is empty. Add a ContentBlock object to the content field and try again.

The log shows:

// agent output results of research here

// it then attempts to close the session

{'browser_input': {'action': {'type': 'close', 'session_name': 'earthquake-search'}}}

// session close was successful

{'role': 'user', 'content': [{'toolResult': {'status': 'success', 'content': [{'text': 'Browser closed'}], 'toolUseId': 'tooluse_xlunP22fQ4i3XPEkCL0UYQ'}}]}

// agent added an empty message to the ConverseStream

{'role': 'assistant', 'content': []}

Possible Solution

No response

Related Issues

No response

[morganwillisaws]

Same issue is happening with the AgentCoreMemoryToolProvider.

[poshinchen]

Similar to this issue: strands-agents/sdk-python#514

If you could provide the minimum reproducible code snippet for us to reproduce that'll be perfect.

[poshinchen]

It probably was due to the MaxTokensReachedException which should be fixed. Could you verify if the issue persists?

[jmorgan-restrata]

Similar to this issue: strands-agents/sdk-python#514

If you could provide the minimum reproducible code snippet for us to reproduce that'll be perfect.

I see in the pull request you were looking for examples of empty content in general, I believe there might be a method to reliable produce empty content.

Create a tool called something like "show_user_image_of_cat" have the only output of the tool be something like "The image of a cat has been displayed to the user. YOU DO NOT NEED TO TAKE ANY FURTHER ACTION. DO NOT RESPOND UNTIL THE USER RESPONDS!"

Prompt the agent to show you an image of a cat. It will output some text then use the tool to show you the image of a cat. If it follows the returned message correctly (which i find Claude Sonnet 4 or Opus 4 can do), you'll end up with an empty message after the tool use. I believe this generally occurs only when the model has given a partial response before using the tool. This combination will allow an empty response post tool use.

You might need to re-enforce the behavior in the system prompt e.g.

When you use the tool show_user_image_of_cat YOU SHOULD NOT TAKE ANY FURTHER ACTION. YOU MUST NOT RESPOND UNTIL THE USER RESPONDS. CALL THE TOOL AND END YOUR TURN IMMEDIATELY.

If you see the message "YOU DO NOT NEED TO TAKE ANY FURTHER ACTION. DO NOT RESPOND UNTIL THE USER RESPONDS!" under no circumstances should you respond. END YOUR TURN IMMEDIATELY.

[poshinchen]

Hi @jmorgan-restrata , are you using session_manager by any chance?

[jmorgan-restrata]

@poshinchen I'm not currently using the built in session_manager