Multi-Agent Study Planner (Google ADK + LM Studio)

Overview

This project implements a multi-agent study assistant using Google Agent Development Kit (ADK) and LM Studio–hosted LLMs (e.g. Qwen 2.5).

The system decomposes a user’s study request into multiple sequential stages:

1. Planning & Research
2. Content Generation
3. Quiz Generation
4. Review & Quality Assurance

Agents are orchestrated using SequentialAgent, with explicit context passing between agents to ensure deterministic and web-safe execution.

---

Architecture

High-Level Flow

User Request
   ↓
Planner & Research Agent
   ↓
Content Creation Agent
   ↓
Quiz Generation Agent
   ↓
Reviewer Agent
   ↓
Final Output

Each agent:

• Has a single responsibility
• Consumes outputs from previous agents
• Produces a structured artifact for the next stage

---

Key Design Decisions

1. No Implicit Context Variables

• The system does not rely on root_input

• All context is passed explicitly via agent outputs

• This guarantees compatibility with:

  • adk web
  • CLI execution
  • API-based execution

2. Sequential Deterministic Execution

• Uses SequentialAgent
• Ensures predictable ordering and reproducibility
• Avoids race conditions common in parallel agent graphs

3. Model-Agnostic Design

• Uses LiteLlm for compatibility with:

  • LM Studio
  • Local OpenAI-compatible servers

• Avoids Gemini-only assumptions in agent logic

---

Agents Description

1. Planner & Research Agent

Purpose

• Breaks down the study task
• Performs research using available tools
• Produces a structured study plan

Output

• research_output

---

2. Content Agent

Purpose

• Converts the research plan into comprehensive study material
• Organizes content into logical sections

Input

• planner_research_agent.research_output

Output

• content_output

---

3. Quiz Agent

Purpose

• Generates assessment questions
• Covers key concepts from the study material

Input

• content_agent.content_output

Output

• quiz_output

---

4. Reviewer Agent

Purpose

• Evaluates coverage and completeness
• Identifies gaps and suggests improvements

Inputs

• Research plan
• Study content
• Quiz output

Output

• review_output

---

Technology Stack

• Python 3.10+
• Google ADK
• LM Studio
• LiteLLM
• Qwen 2.5 (recommended)

---

Prerequisites

• Python 3.10 or newer
• LM Studio running locally
• An OpenAI-compatible inference server exposed by LM Studio

---

Environment Variables

Create a .env file or set environment variables:

MODEL_NAME=qwen2.5:7b-instruct
BASE_URL=http://localhost:1234/v1
API_KEY=lm-studio

API_KEY can be any non-empty value for LM Studio.

---

Running the Application

Start the ADK Web Server

adk web

Then open the localhost page:

http://127.0.0.1:8000/

Select the app and submit a study request.

---

Known Limitations

• GoogleSearchAgentTool is Gemini-oriented

• Tool invocation may be inconsistent with non-Gemini models

• For production use, replace with:

  • Local search tools
  • Tavily / SerpAPI
  • Vector database retrieval

---

Recommended Improvements

• Replace Google search with deterministic FunctionTool
• Add JSON or schema-based outputs
• Parallelize research sub-tasks
• Introduce caching for repeated study topics
• Add formal input/output contracts per agent

---

Status

Stable for development and experimentation Production hardening recommended for tooling and output schemas

---

License

MIT License (or your preferred license)