This example demonstrates the use of Amazon Bedrock Agents multi-agent collaboration with its built-in Routing Classifier feature. By simply enabling that mode for your supervisor, Bedrock automatically routes to the correct collaborating sub-agent using LLM intent classification optimized to route with sub-second latency. Contrast that with a traditional supervisor that must go through its own orchestration loop, a more expensive proposition that can take 3-6 seconds depending on which LLM you are using. This feature is most valuable when trying to build a unified customer experience across a set of sub-agents. In our example, we have 3 collaborators: one for general mortgage questions, one for handling conversations about Existing mortgages, and another for dealing with New mortgages.
Try sample prompts:
- "what's my balance?"
- "what are rates looking like over the past 2 weeks for new mortgages?"
- "what's your general guidance about refinancing mortgages?"
Each of these gets routed quickly to the right sub-agent for subsequent processing. Conversation switching is seamless.
Routing also reverts automatically to full Supervisor mode if the request is truly not mapping cleanly to a single collaborator. For example:
- "what interest rate do i have on my existing mortgage, and how does it compare to rates for new mortgages"
- Clone and install repository
git clone https://github.com/awslabs/amazon-bedrock-agent-samples
cd amazon-bedrock-agent-samples
python3 -m venv .venv
source .venv/bin/activate
pip3 install -r src/requirements.txt- Deploy Amazon Bedrock Agents
python3 examples/amazon-bedrock-multi-agent-collaboration/mortgage_assistant/main.py \
--recreate_agents "true"- Invoke
python3 examples/amazon-bedrock-multi-agent-collaboration/mortgage_assistant/main.py \
--recreate_agents "false" \
--team_name "New England Patriots"- Cleanup
python3 examples/amazon-bedrock-multi-agent-collaboration/mortgage_assistant/main.py \
--clean_up "true"This project is licensed under the Apache-2.0 License.