Multi-Agent Workflow for RFP Challenges

This project showcases a multi-agent system where two agents collaborate to address RFP (Request for Proposal) challenges:

• Client Agent: Initiates the RFP process by presenting the proposal.
• Supplier Agent: Engages with the client to refine and validate the RFP, ensuring clarity and completeness.

The workflow is designed to emulate real-world scenarios where agents negotiate and validate proposals, fostering better communication and understanding between stakeholders.

Based on: https://www.diagrid.io/blog/building-effective-dapr-agents

Prerequisites

• Python 3. (recommended: Python 3.10 or 3.11)
• pip (Python package manager)
• python3.10-venv
• Dapr CLI (Install instructions)
• Docker (for Dapr state store/pubsub components)
• Redis (for Dapr state store/pubsub, can be run via Docker)
• Azure OpenAI resource (for LLM-powered agents)
• Linux or WSL2 (recommended for multi-app Dapr YAML support)

Prepare the environment

Note:
These steps are intended to be run on a Linux environment. The command dapr run -f <yaml> requires Linux to work properly with multiple applications defined in a YAML file.
If you are on Windows, please install and run everything inside Ubuntu (WSL) or another Linux environment.

cd src
python3 -m venv .venv
source .venv/bin/activate
dapr init
pip install -r requirements.txt

Tip:
To exit the Python virtual environment (.venv), simply run:

deactivate

Create a .env file with the following content:

AZURE_OPENAI_API_KEY="<your-azure-openai-api-key>"
AZURE_OPENAI_ENDPOINT="https://<your-azure-openai>.openai.azure.com/"
AZURE_OPENAI_DEPLOYMENT="<openai-deployment-name>"
AZURE_OPENAI_API_VERSION="2025-01-01-preview"
DAPR_LLM_COMPONENT_DEFAULT="openai"

Run the RPF Challenge Workflow

dapr run -f dapr-multi-agent-rpf-challenge-workflow.yaml

To monitor Dapr traces and metrics, you can open Zipkin in your browser. Default Zipkin URL: http://localhost:9411

Workflow Overview

Reviewing the Conversation with HumanResponse

After the workflow completes LLMOrchestrator_state.json, you can use the HumanResponse program to review the full conversation between the agents in a human-friendly Markdown format, and generate an AI-powered summary.

What does HumanResponse do?

• Reads the conversation state from the workflow.
• Generates a Markdown file (conversation.md) with the full dialogue between the ClientAgent and SupplierAgent.
• Optionally, creates a summarized version of the conversation using Azure OpenAI, saved as conversation_resume.md.

How to use it:

cd poc/humanresponse
python3 app.py conversation.md

This will generate a summary file named conversation_resume.md in the same directory, containing the main points and agreements from the conversation.

---

Troubleshooting

If you encounter problems with Dapr (for example, errors with state stores, pubsub, or sidecars not starting), you can reset your Dapr installation with:

dapr uninstall --all
dapr init

This will remove all Dapr components and re-initialize Dapr.

---

Credits

The files and the approach to responding to RFPs were inspired by examples from Responsive.io.