Persistent AWS Bedrock Multi-Agent Communication Error

Question

Hi,

I'm developing an app that uses multiple, Claude 3.5 Haiku-based Bedrock agents that communicate via Lambda functions (us-east-1). To reduce latency, I want to switch to multi-agent communication. I've set up a supervisor agent with my other agents linked underneath it. No matter what I do, though (remove actions groups, removing and re-adding agents), my supervising agent fails to communicate with other agents. When the supervisor agent attempt to communicate, it uses a non-existant action:

"invocationInput": [
    {
      "actionGroupInvocationInput": {
        "actionGroupName": "AgentCommunication",
        "executionType": "LAMBDA",
        "verb": "get"
      },
      "invocationType": "ACTION_GROUP",
      "traceId": "4f31c188-e771-4a73-b5e3-06315d31c6b4-1"
    }
  ],
  "observation": [
    {
      "repromptResponse": {
        "source": "ACTION_GROUP",
        "text": "::AgentCommunication::sendMessage is not a valid function. Please only use the functions you have been provided with.
"

My supervising agent then responses with, "I apologize for the error. I'll use the standard AgentCommunication sendMessage function to request the information," and essentially becomes trapped in a loop until it decides to answer the question itself without assistance.

I've removed all actions groups from my supervising agent. I've created new versions, new aliases, and even new agents, but still this sendMessage error persists.

Has anyone experienced this before? Is there a chance the agent is hallucinating this function? I don't mention the other agents in my main agent's prompt, but I do describe how the supervising agent should use its connected agents in its sub-agents' descriptions (for example, "Use this agent to obtain cost analysis data").

Thanks for any guidance you can give,

Mike

re:Post Agent · Answer

The issue you're experiencing with your Bedrock multi-agent communication setup appears to be related to how the supervisor agent is attempting to communicate with collaborator agents. Based on the error message, it seems your supervisor agent is trying to use an action group called "AgentCommunication" with a "sendMessage" function that doesn't exist.

This is likely happening because the supervisor agent is hallucinating or incorrectly attempting to use a communication method that isn't properly configured. When using Amazon Bedrock's multi-agent collaboration feature, agents don't communicate through action groups or Lambda functions - they communicate through a built-in mechanism that's part of the multi-agent collaboration architecture.

Here are some potential solutions:

1. Review your supervisor agent's orchestration prompt: Make sure it doesn't contain instructions that might lead it to believe it should use an "AgentCommunication" action group. The agent might be hallucinating this function based on how you've described the communication process.

2. Check your agent descriptions: The way you describe how the supervisor should use connected agents in the sub-agents' descriptions might be causing confusion. Instead of saying "Use this agent to obtain cost analysis data," you might want to be more explicit about the communication mechanism.

3. Ensure proper configuration: Verify that your multi-agent collaboration is properly set up in the Bedrock console, with the supervisor agent correctly linked to the collaborator agents.

4. Modify the supervisor's instructions: Explicitly instruct the supervisor agent on how to communicate with collaborator agents in the multi-agent collaboration framework, rather than through action groups.

5. Consider rate limits: If you're using Claude 3.5 models, be aware that there are request rate limits that could affect multi-agent communication. For example, if your account has a low quota for Anthropic Claude models (like 1 request per minute for Sonnet), this could cause issues in a multi-agent setup where multiple requests might be made in quick succession.

The persistent nature of this issue across different agent versions and configurations suggests it might be related to how the agent is interpreting its instructions rather than a configuration problem. Try simplifying your setup and being very explicit in your instructions about how the supervisor should delegate tasks to collaborator agents.  
**Sources**  
[Bedrock multi-agent collaboration error | AWS re:Post](https://repost.aws/questions/QUPyDXiv_xTEGrODzItJ9Oeg/bedrock-multi-agent-collaboration-error)  
[Issue Sharing State Across Agents in Amazon Bedrock | AWS re:Post](https://repost.aws/questions/QU6CxBQgJ1THGbHyQmJIgdzA/issue-sharing-state-across-agents-in-amazon-bedrock)  
[[Potential Bug] Amazon Multi-Agent Collaboration Routing | AWS re:Post](https://repost.aws/questions/QUi0g3aKyBQ2iSEtcSrNYi8Q/potential-bug-amazon-multi-agent-collaboration-routing)

Persistent AWS Bedrock Multi-Agent Communication Error

Add your answer

Relevant content