Foundry Responses hosted MCP calls run sequentially despite parallel_tool_calls=true

[Ritesh Srivastava (riteshksriv)]

Technical Feedback

Describe the precise feedback.
Azure AI Foundry Responses/Hosted Agents accepts parallel_tool_calls: true and emits two independent hosted MCP calls, but executes them sequentially rather than concurrently.

Environment:

• Azure AI Foundry project in East US 2
• Hosted agent model: gpt-5.4-mini
• Public MCP server: https://learn.microsoft.com/api/mcp
• Allowed tool: microsoft_docs_search
• require_approval: "never"
• Related Agent Framework report: .NET: [Bug]: Hosted MCP calls remain sequential despite parallel_tool_calls=true microsoft/agent-framework#7825

A live .NET request capture immediately before transport proves the request reaches Foundry with the intended settings:

{
  "parallel_tool_calls": true,
  "tools": [{
    "type": "mcp",
    "server_label": "microsoft_learn",
    "server_url": "https://learn.microsoft.com/api/mcp",
    "allowed_tools": { "tool_names": ["microsoft_docs_search"] },
    "require_approval": "never"
  }]
}

The prompt explicitly required exactly two independent documentation searches to be issued together before waiting for either result. In three runs, Foundry generated two distinct MCP calls, but the second call's in_progress event occurred only after the first call's completed event:

1. Response caresp_fac59ca04cf8dc2600D48D83eN4Mxud4XFpUtwfVcaMeXLbueZ: call 1 completed at 5547 ms; call 2 started at 5570 ms (23 ms later).
2. Response caresp_8fdf2031a7755a5c00jjqh4OL3UBu29ahT2cVQvGNAKczLypxy: call 1 completed at 3669 ms; call 2 started at 3674 ms (5 ms later).
3. Response caresp_44824cef832e1c6900bHUN5qycxeNruKmI7Z5hpyCcl57gEulK: call 1 completed at 3340 ms; call 2 started at 3349 ms (9 ms later).

There was no overlap in any run. The same behavior was observed through a deployed Python Hosted Agent configured directly with parallel_tool_calls=True, ruling out a .NET-only client serialization issue.

Desired Outcome

Describe the desired outcome.
When a Responses request has parallel_tool_calls: true and the model emits multiple independent hosted MCP calls in one turn, Foundry hosted-tools orchestration should start and execute those calls concurrently. If hosted MCP concurrency is not supported, the service should document that limitation and clarify the scope of parallel_tool_calls for hosted tools.

Current Workaround

Describe the current workaround.
No equivalent hosted-agent workaround is available. A client-controlled agent can invoke MCP tools locally and schedule independent calls concurrently, but that gives up Foundry-hosted MCP execution. Prompts that explicitly request one tool-call batch still produce sequential hosted MCP lifecycle events.