Can I make the framework to return the tool result directly, without rewriting or summarizing it via another LLM pass？

[eison (dlwangac)]

I need the agent only do tool-routing without a second LLM call for final synthesis.

Current behavior

When tools are enabled, the framework executes tool calls and then performs another LLM step to produce the final assistant response.
In some scenarios, we need the framework to return the tool result directly, without rewriting or summarizing it via another LLM pass.

Desired behavior

• The model can still decide which tool to call (routing).
• After tool execution, the framework can short-circuit and return the raw tool output as the final response.
• No second LLM completion is made for synthesis/rephrasing.

Is there any option setting or other method to achieve this?

Thank you!

[Evan Mattson (moonbox3)]

Hi eison (@dlwangac), yes, this is done via middleware. Are you using MAF Python or C#? I can provide some samples depending on which language you're using.

[eison (dlwangac)]

Hi Evan Mattson (@moonbox3) ,thank you very much.
I'm using MAF Python.

[Evan Mattson (moonbox3)]

Please have a look at this sample: https://github.com/microsoft/agent-framework/blob/main/python/samples/02-agents/middleware/override_result_with_middleware.py

[Evan Mattson (moonbox3)]

eison (@dlwangac) you can also read about middleware in our docs here: https://learn.microsoft.com/en-us/agent-framework/agents/middleware/?pivots=programming-language-python

[eison (dlwangac)]

Hi Evan Mattson (@moonbox3)

This approach doesn't seem to prevent the second LLM call;
it merely overwrites the result of the second LLM call.

Is there a way to prevent the framework from making a second LLM call?

[Evan Mattson (moonbox3)]

You need to make sure you're terminating, so the call doesn't go back to the LLM. Middleware allows you to get a hold of the function result, and optionally returning early, before the model has a chance to respond again.

The normal flow for function calling is:

User input -> model call -> model wants a tool call -> tool runs -> model sees tool result -> model final output based on tool result

Using middleware:

User input -> model call -> model wants a tool call -> tool runs -> exit function calling loop via middleware -> result contains tool result (not modified via model)

[eison (dlwangac)]

Hi Evan Mattson (@moonbox3)

I see.
I need to prevent LLM calls from being made by controlling whether [await call_next()] is executed.

Thank you very much.