Standalone Multi-Agent AI Core Engine

A framework-style, standalone multi-agent AI system designed to demonstrate how multiple specialized agents can collaboratively solve a user query using planning, explanation, evaluation, and feedback.

This project focuses purely on agent orchestration and communication, without UI or product-level complexity, making it ideal for understanding and explaining multi-agent AI architectures.

Project Overview

Modern AI systems require more than a single model response.
This project implements a clean, minimal multi-agent pipeline where each agent performs a single responsibility and is coordinated by a central orchestrator.

The system processes a user query through four agents:

1. Planner Agent – breaks the task into logical steps
2. Explainer Agent – generates an initial structured answer
3. Evaluator Agent – critiques correctness and identifies gaps
4. Feedback Agent – refines and improves the final response

This creates a closed feedback loop, enabling self-improving AI behavior.

System Architecture

User Query
↓
Orchestrator
↓
Planner Agent → Task Decomposition
↓
Explainer Agent → Draft Answer
↓
Evaluator Agent → Quality Check
↓
Feedback Agent → Improved Final Answer

Folder Structure

multi_agent_core/
│
├── agents/
│ ├── base_agent.py # Abstract base class for all agents
│ ├── planner.py # Task decomposition agent
│ ├── explainer.py # Answer generation agent
│ ├── evaluator.py # Answer evaluation agent
│ └── feedback.py # Answer refinement agent
│
├── orchestrator.py # Controls agent execution order
├── llm_client.py # Centralized LLM abstraction (mocked)
├── main.py # Entry point
└── README.md

Agents Description

Planner Agent

• Breaks a user query into clear, logical steps
• Normalizes LLM output into a structured list
• Prevents prompt leakage to downstream agents

Explainer Agent

• Generates a structured explanation
• Follows the plan produced by the Planner
• Focuses on clarity and completeness

Evaluator Agent

• Critiques the generated answer
• Identifies missing points and issues
• Produces structured evaluation output

Feedback Agent

• Uses evaluator feedback
• Improves clarity, correctness, and coverage
• Closes the multi-agent feedback loop

LLM Abstraction

All model interactions are routed through a single LLMClient.

This design allows easy replacement with:

• OpenAI
• Gemini
• Local LLMs

without changing agent logic.

How to Run

python main.py

Example query (inside main.py):

"Explain Gradient Descent"