Hey, I tried using this WIP PR, but found I couldn't run agents concurrently. Maybe it's because self.client.converse_stream(**request) in BedrockModel.stream() is a blocking boto3 call that prevents asyncio.gather() from running multiple agents in parallel?

Hey, I tried using this WIP PR, but found I couldn't run agents concurrently. Maybe it's because self.client.converse_stream(**request) in BedrockModel.stream() is a blocking boto3 call that prevents asyncio.gather() from running multiple agents in parallel?

@pgrayy this might be something to explore/file an issue for. We've seen something similar before strands-agents/tools#91 (comment). Ultimately this might accelerate the usage of the experimental client.

Hey, I tried using this WIP PR, but found I couldn't run agents concurrently. Maybe it's because self.client.converse_stream(**request) in BedrockModel.stream() is a blocking boto3 call that prevents asyncio.gather() from running multiple agents in parallel?

Yes this is correct. Bedrock does have an async client in the works though (source). We will likely work out a process to give customers an option to opt-in for its use while we wait for a 1.0 release. In the meanwhile, we will update our docs to identify current limitations with async. I will note though that for strands 1.0, some model providers will be converted to full async. For example, the OpenAIModel provider will be updated to use AsyncOpenAI.

Ah okay thanks, didn't know about the experimental async client. Providing that option would be great!

For context, my team uses Strands heavily and needs offline concurrent utilities for evaluation:

• Running all agents × all test queries (many-to-many)
• Running a single model.structured_output against multiple llm-as-a-judge prompts (one-to-many)