Multi-Agent Marketing Workflow: Collaborative AI Pipeline Based on OpenAI Agents SDK

This article introduces the open-source project "multi-agent-workflows" published by developer liannfinlayson26 on GitHub. Built using the OpenAI Agents SDK, this project creates a collaborative AI pipeline where three specialized agents—Creative Director, Strategist, and Copywriter—work together to complete the full marketing campaign workflow from market research to tweet generation. The project's core philosophy is "divide and conquer", allowing each agent to focus on its area of expertise, resulting in collaborative effects that surpass a single AI.

As the capabilities of large language models improve, a single AI assistant can hardly meet complex business needs. Multi-agent systems have become a new paradigm through specialized AI collaboration. The goal of this project is to simulate the collaboration of a real marketing team: after the user provides a campaign theme, the system automatically coordinates the three agents to complete the entire workflow. The data flow of the workflow is: brief → Creative Director → Strategist → Copywriter → tweets, drawing on the traditional marketing team model and combining the advantages of intuitiveness and automation.

1. Creative Director: Receives the theme, conducts market research using Tavily search, and generates 3-5 creative proposals (based on real network information to improve feasibility); 2. Strategist: Evaluates the creative proposals, selects the top 2, and provides reasons for recommendation (enhancing decision interpretability); 3. Copywriter: Converts the selected proposals into tweets that align with the brand tone and automatically adds hashtags (directly deployable to social media).

1. Custom Function Tools: Encapsulate Tavily search with @function_tool, allowing agents to independently call and expand their capabilities; 2. Structured Output: Each agent uses Pydantic models to define output formats, ensuring structured and predictable data transmission; 3. Multi-Agent Orchestration: The manager_run() function calls agents in sequence and passes data, maintaining a shared SQLiteSession; 4. Shared Memory: SQLiteSession stores conversation history, allowing all agents to access the complete context and ensuring workflow coherence.

Tech Stack: OpenAI Agents SDK (core), Python, Pydantic, Tavily, gpt-5-mini, Jupyter Notebook. Quick Start: 1. Create a virtual environment and install dependencies from requirements.txt; 2. Copy .env.example to .env and fill in OpenAI and Tavily API keys; 3. Open multi_agent_workflows.ipynb, run cells in order, and call manager_run() to execute the full workflow (example: Launch eco-friendly water bottle campaign in Bali).

Cost Control: It is recommended to set consumption limits for OpenAI/Tavily. Optimization strategies include using gpt-5-mini, caching search results, using simulated data in non-production environments, and limiting max_tokens. Expansion Possibilities: Add Visual Designer/Data Analyst agents; introduce parallel execution; add human-machine collaboration nodes; support more output formats like blogs/emails; integrate image generation/social media publishing tools.

Value: Specialized division of labor (better results than a single agent), interpretability (clear step outputs), reusability (agents can be reused across workflows), scalability (easy to add new agents). Limitations: Increased latency (sequential execution), accumulated costs (multiple API calls), error propagation (upstream errors affect downstream), coordination complexity (increased difficulty in state management).

This project is an elegantly designed and simply implemented multi-agent example, covering core concepts such as tool definition, structured output, and agent orchestration, providing an entry-level case for developers. As AI evolves, multi-agent systems will play a role in more fields, and this project demonstrates the future scenario of specialized AI collaboration to complete complex tasks.