MCP tool call returns "Not a valid request: <uuid>" or empty success —in AI Agent Studio

[tamilb]

Hi all,

I'm running into inconsistent behavior with MCP tool responses in AI Agent Studio and can't identify a clear pattern. Hoping someone from the community or product team can help.

Setup:

• ServiceNow is acting as the MCP client.
• External MCP server registered in sn_mcp_server (using Streamable HTTP transport).
• Tool returns a JSON response (verified correct on the MCP server side every time).

Issue:

I'm seeing two different failure modes when calling tools, with no clear pattern.

Case 1 — Empty response with success status:

The tool returns valid JSON on my MCP server, but in sn_mcp_execution_logs the response field shows up empty:

{
"tool_execution_id": "4400fbb82bec83502cebfccdc891klds",
"response": "",
"status": "success",
"error": null
}

Case 2 — "Not a valid request" message:

After modifying the tool response, the same call returns:

{
"tool_execution_id": "0778fbb82bec67502cebfdddc891bf02",
"response": "Not a valid request: f408b81f-3ee2-48a2-8dd8-93c2218afabe",
"status": "success",
"error": null
}

Case 3 — Works correctly:

When I reduce the size of the tool's response payload, ServiceNow displays the actual response correctly with no other changes.

Key observations:

1. My MCP server returns the correct tool result every single time (verified in server logs).
2. In all failing cases, status is "success" and error is null — but the response field is either empty or contains the "Not a valid request" string instead of the actual output.
3. The size of the tool response seems related (smaller responses succeed, larger ones fail), but I can't pin down a consistent threshold or pattern.
4. The same tool call with identical arguments behaves differently depending on response size.

Questions:

1. Has anyone else seen this empty-response or "Not a valid request" behavior with MCP tools in AI Agent Studio?
2. Is there a known response size limit on the MCP client side?
3. Is there a configurable property to adjust this limit?
4. Is the recommended pattern to design tools with pagination/smaller payloads, or is there a fix/workaround?

Any insights, documentation, or shared experiences would be greatly appreciated.

Thanks!