Add Target for AgentCore Runtime #113
Description
Today I had a quick go at sketching out a Target implementation for agents deployed on Bedrock AgentCore Runtime.
It seems generally pretty feasible within the current patterns... But I think the main design decision/friction point will be how to handle the facts that AgentCore supports arbitrary request & response payloads, and expects response streaming?
In SageMakerEndpointTarget, this was handled by configuring:
- A template request payload JSON
- A JSONPath to overwrite in the request payload with the new prompt
- A JSONPath to query from the response (also assumed JSON) to extract the reply text.
The AgentCore docs propose a generic code sample here for handling both application/json and text/event-stream response streams.
As far as I'm aware, agent-evaluation doesn't have a general solution for introducing multi-modal tests yet?
So, any opinions on how to set up configuration to make an AgentCore target suitably flexible but easy-to-use?