[BUG] Do not invoke synchronous `call()` on LLM from asynchronous workflow in `_export_output` / converter

[andrewisplinghoff]

Description

In an asynchronous workflow started via akickoff(), there should not be synchronous calls being made as that blocks the event loop. If the output of a task requires converting to fit the desired output format, there are currently synchronous calls made to the LLM via several invocations of self.llm.call() here in the converter module (https://github.com/crewAIInc/crewAI/blob/c14abf1758dd3aafc57ad7f17569174c7cc1ea68/lib/crewai/src/crewai/utilities/converter.py), which blocks the event loop.

Calls from async methods here:

crewAI/lib/crewai/src/crewai/task.py

Line 605 in c14abf1

pydantic_output, json_output = self._export_output(result)

crewAI/lib/crewai/src/crewai/task.py

Line 1258 in c14abf1

pydantic_output, json_output = self._export_output(

_export_output then delegates to the converter module here:

crewAI/lib/crewai/src/crewai/task.py

Line 1023 in c14abf1

model_output = convert_to_model(

Steps to Reproduce

Use some example workflow that uses a Task(output_json=...)orTask(output_pydantic=...) to require a specific output format, e.g.:
https://github.com/crewAIInc/crewAI-examples/blob/b4c0ef6522c37d4148f8372a731fd2851fdcb864/flows/self_evaluation_loop_flow/src/self_evaluation_loop_flow/crews/x_post_review_crew/x_post_review_crew.py#L32

Invoke the workflow asynchronously using akickoff().

Expected behavior

Only calls to await llm.acall() should be made from an asynchronous workflow.

Screenshots/Code snippets

Operating System

Debian 12 Bookworm

Python Version

3.12

crewAI Version

1.12.2

crewAI Tools Version

Virtual Environment

Venv

Evidence

Possible Solution

Additional context

[zday123]

Also experiencing extremely degraded throughput in async kickoff calls. In versions prior to 1.0.0, I could launch a few hundred coroutines and they would be processed within a minute. Now, even with akickoff() (changed from kickoff_async()), I am not getting nearly the same throughput.

[shivam2407]

I'd like to work on this. The root cause is that _export_output() calls convert_to_model() which eventually calls the synchronous self.llm.call() and instructor client from within async methods (_aexecute_core, _ainvoke_guardrail_function). My plan is to add async variants throughout the chain: _aexport_output(), aconvert_to_model(), Converter.ato_pydantic()/ato_json(), and InternalInstructor.ato_pydantic()/ato_json() — all using llm.acall() and async instructor clients. Sync paths remain untouched for backward compatibility.