Implementing Multi-Agent Architecture Using Amazon Bedrock Models for Supervisor-Controlled Decision Making and Retry Mechanism

[VinuthaRaghavendra]

Hi everyone,

I'm developing a multi-agent system for an automation project using Amazon Bedrock models, and I'm exploring ways to implement an architecture with the following agents:

Supervisor - Manages and coordinates the sequence of actions.
Action Agent - Executes specified actions using a Bedrock model.
Verification Agent - Uses a Bedrock model to verify action outcomes.
Retry Agent - Uses Bedrock model-based logic to determine if retries are needed.
Requirements:
The Supervisor agent will receive a sequence of actions to execute and needs to:

Process each action individually and dynamically determine which agent should handle the action (either Action agent, Verification, or Retry).
Capture the responses from each agent and, based on the response, decide if a retry is necessary before moving to the next action.
Continue executing the sequence of actions while ensuring each step is verified and retried if needed.

Challenges:
Structuring the Supervisor's decision-making to select the right agent using Bedrock models.
Coordinating responses across Bedrock model-powered agents, especially handling retry logic seamlessly.
Capturing and tracking each agent’s responses so that the Supervisor can make informed decisions on retries or progression.

1. What would be an effective way to structure the Supervisor’s decision-making logic with Bedrock models to dynamically select the appropriate agent for each action? and also code snippet for it
2. Are there best practices for coordinating Bedrock model-based responses across agents, especially when retry logic is required?
3. How can I capture and track responses in a structured way to allow the Supervisor to decide whether to retry or continue with the next action in the sequence?
4. Any advice, architectural patterns, or similar experiences with Bedrock models would be greatly appreciated!

Thank you!