Execution and Streaming¶

Every provider exposes run, run_async, and stream. The methods preserve provider-native return types and streaming conventions.

Return types¶

Provider	`run`	`run_async`	`stream`
OpenAI	`RunResult`	`RunResult`	`RunResultStreaming`
Gemini	Final `Event`	Final `Event`	`AsyncIterator[Event]`
Grok	`Response`	`Response`	`AsyncIterator[tuple[Response, Chunk]]`
Claude	`BetaMessage`	`BetaMessage`	`BetaAsyncStreamingToolRunner[object]`
Mistral	`ChatCompletionResponse`	`ChatCompletionResponse`	`Iterator[CompletionEvent]`

Asynchronous execution¶

import asyncio

from minions.gpt import ResearchMinion


async def main( ) -> None:
    minion = ResearchMinion(
        model='gpt-5.6-sol',
        instructions='Research the subject and return a concise summary.',
    )
    result = await minion.run_async( 'Research the requested subject.' )
    print( result.final_output )


if __name__ == '__main__':
    asyncio.run( main( ) )

Grok and Mistral accept synchronous or asynchronous local functions during async execution. Synchronous functions run in worker threads.

OpenAI streaming¶

stream = minion.stream( 'Explain the provider boundary.' )

async for event in stream.stream_events( ):
    print( event )

Gemini streaming¶

async for event in minion.stream( 'Research the requested subject.' ):
    print( event )

print( minion.result )

Grok streaming¶

async for response, chunk in minion.stream( 'Research the requested subject.' ):
    print( chunk )

print( minion.result )

Claude streaming¶

runner = minion.stream( 'Research the requested subject.' )

runner is Anthropic's asynchronous streaming tool runner. Consume it through the streaming interfaces provided by the installed Anthropic SDK; it continues through hosted and local tool calls automatically.

Mistral streaming¶

for event in minion.stream( 'Research the requested subject.' ):
    print( event.data )

Result retention¶

After execution, minion.result refers to the most recent provider-native result or streaming object. Gemini also retains invocation events in minion.events.

Limits and failures¶

max_turns must be at least 1. A Grok or Mistral workflow raises RuntimeError if local tool calls exhaust that limit without producing a final answer. Empty required strings, mismatched local tool mappings, and missing applicable credentials fail before useful provider work begins.