multi-agent-AI-system

Multi-Agent AI Marketing Campaign System

An automated, cooperative multi-agent system designed to turn raw product concepts into highly targeted marketing assets using real-time search and an iterative editorial loop.

Overview

The system automates end-to-end campaign asset creation, specifically transforming a raw, high-level product description into a structured market research brief, LinkedIn post, and email campaign.

Why Multi-Agent ?

• Specialized Division of Labor: It splits a complex, multi-step job into highly focused, specialized roles (researching, writing, and editing).
• Adversarial Feedback Loop: It utilizes a Writer vs. Critic dynamic to guarantee quality, closely mimicking a real-world agency where creative output must meet strict editorial standards before publication.

Agent Architecture & Roles

Agent	Core Role	Inputs	Outputs	Specialized Tools
1. Researcher	Uncovers market facts, target audience insights, and competitor details.	Raw product description	JSON containing: target_audience, pain_points (list), marketing_angle	tavily_search_tool: A custom Python web search tool.
2. Writer	Drafts highly engaging social copy and structured emails.	Researcher's JSON brief AND any active feedback from the Critic	JSON containing: linkedin_post, email_copy	None (creative reasoning)
3. Critic	Acts as a strict chief editor to evaluate copy against the research brief.	Researcher's JSON brief AND the Writer's latest draft	JSON containing: approved (boolean), feedback (string)	None (strict evaluation)

---

Agent Communication & Data Flow

Because this pipeline is orchestrated via standard Python code, agents communicate by passing structured JSON strings back and forth using standard JSON schemas at every hand-off.

How Information Flows:

1. The Research Phase: The Researcher returns a JSON block. Python parses this string using json.loads().
2. The Creative Phase: The parsed dictionary is converted back into a formatted string using json.dumps() and injected directly into the Writer's prompt context alongside the current critique string.
3. The Editorial Phase: The Writer's generated JSON response is similarly loaded and dumped directly into the Critic's evaluation prompt.
4. The Decision Loop: The Critic's JSON response is parsed to inspect the "approved" boolean:
   • If False: The "feedback" string is stored, updating the critic_feedback variable, and the loop triggers the Writer for a revised draft.
   • If True: The loop terminates and the final assets are approved.

Frameworks and tools :

• LLM Engine: gemini-2.5-flash (chosen for its exceptionally fast response times, high context window, low cost, and reliable structural JSON adherence).
• AI Orchestration: Google GenAI SDK (utilizing the modern google.genai client and types.GenerateContentConfig for tool definitions).
• Search Engine Tool: Tavily API (via TavilyClient) to query the web and gather clean, parsed, developer-friendly Markdown content rather than messy raw HTML.
• **Environment Management:
• Terminal UI: rich to display colored and formatted JSON print outputs natively in the CLI.

Getting Started & How to Run

1. Install Dependencies

pip install google-genai tavily-client python-dotenv rich Set Up Environment Variables Create a file named .env in the same directory as the script and insert the API keys:

GEMINI_API_KEY=XXXXXXXXXXXXXXXX TAVILY_API_KEY=XXXXXXXXXXXXXXXX

then run : "python agent.py"

Challenges :

Lack of Conversational Context: The current model lacks a rolling conversation thread of previous messages, making it harder for the model to see the step-by-step progression of drafts and feedback over multiple loops.

Tool Execution Failures (tavily_search_tool): If Tavily's API keys expire or their servers face downtime, the function catches the error and returns "Error executing search tool: ...". Gemini doesn't realize this is a system crash; it treats the error string as factual search context, leading to poor, confused marketing briefs.

JSON Formatting & Parsing Failures: Relying on raw string conversions can occasionally cause code execution breaks if the model returns malformed JSON brackets.

if i had more time i would try to use frameworks like LangGraph , Use Pydantic to enforce schemas and fix the JSON format failure problem ,Provide the full iteration history so the model sees the progression of drafts and feedback.

DEMO :

input : "An AI-powered calendar app that automatically declines low-priority meetings for software engineers based in 2026."