How MultiAgentPPT Generates Slides with AI Agents: Architecture and Code Walkthrough

MultiAgentPPT Implementation Overview

The MultiAgentPPT project (https://github.com/johnson7788/MultiAgentPPT) implements an A2A+MCP+ADK multi‑agent system that streams concurrent PPT generation, inspired by https://github.com/allweonedev/presentation-ai.

The workflow is visualized in two diagrams (shown below) and consists of four agents.

Core Agents

Outline Generation Agent creates an initial content outline based on user requirements.

Topic Splitting Agent divides the outline into multiple research topics.

Research Agent runs in parallel, with each sub‑agent performing deep research on an assigned topic.

Summary Agent aggregates the research results and streams the final PPT content back to the frontend.

The backend implementation resides at https://github.com/johnson7788/MultiAgentPPT/tree/main/backend.

Data Source and Retrieval

The system relies on a Retrieval‑Augmented Generation (RAG) pipeline, requiring a knowledge base that is kept up‑to‑date via a web crawler. The crawler code (

weixin_search.py

) searches WeChat public accounts, obtains real URLs, and extracts article content.

The extracted WeChat content is then fed into the knowledge base for RAG.

Sub‑Agent Execution Logic

Key modules include

simpleOutline

(frontend outline test),

simplePPT

(simple PPT test),

slide_outline

(outline generation with retrieval), and

slide_agent

(PPT generation from outline).

slide_outline

uses RAG to fetch relevant articles, assembles them into a prompt, and sends it to the LLM for generation.

The

slide_agent

further splits into components:

research_topic

,

split_topic

, and

summary_writer

, each driven by specific prompts.

The

split_topic

prompt parses the outline into independent research topics, while

research_topic

creates parallel researchers to gather material. The

SummaryAgent

compiles the results using XML‑style prompts.

References

1. https://github.com/johnson7788/MultiAgentPPT