Madison: Open Source AI Marketing Framework | Join the Project

Building a full marketing department's worth of research, content, and optimization work usually takes a team of specialists and months of lead time. Madison is an attempt to compress that into a coordinated system of AI agents, and it's structured as an open-source research project that anyone interested in agentic AI, marketing, or both can join and help build.

Five specialized agent layers

Madison is organized around five distinct layers, each responsible for a different piece of the marketing problem. Intelligence agents monitor brand reputation across thousands of sources simultaneously, tracking sentiment and market dynamics at a pace no human team could match. Research agents analyze data from thousands of customers to identify market segments and build detailed customer profiles. Content agents generate personalized marketing materials for many channels at once, aiming for consistent brand messaging across platforms without the weeks of manual work that usually takes. Experience agents go beyond analysis into service, building digital concierges and personalized recommendations. Performance agents continuously test and optimize marketing decisions using mathematical algorithms rather than guesswork.

What each layer is actually good for learning

Each layer maps to a different technical interest, which is part of what makes Madison useful as a learning project rather than just a product to use. The content and research agents are a natural fit for anyone interested in combining agentic AI with generative AI, since the challenge there is coordinating generation with real customer data rather than generating in a vacuum. The research agents, which create synthetic customer personas from survey data, are a good fit for anyone interested in verifying truth in AI, since a synthetic persona is only useful if it's an honest reflection of real customer patterns. The performance agents, which rely on continuous testing and optimization, are built for people interested in multi-armed bandits and reinforcement learning specifically.

The orchestration layer as the real ambition

The most ambitious piece of Madison is its orchestration layer, the part meant to coordinate all five agent types seamlessly, the way a conductor coordinates a symphony orchestra rather than a set of instruments playing independently. This is described as the heart of what makes the framework agentic rather than just a collection of separate tools: the potential for the system to check itself for errors and learn from results, so that a marketing investment works harder and smarter over time instead of running the same static campaign indefinitely.

An open invitation, not a finished product

Madison is explicitly framed as unfinished and open. It's a research project, not a polished SaaS tool, and its value to a contributor comes from getting to help shape a five-layer agentic system from a meaningfully early stage. The framework is described as adaptable across industries, whether manufacturing, retail, or something else entirely, provided someone customizes it for that specific need, which is itself part of the open work still to be done.

Key takeaways

• Madison is an open-source agentic marketing framework built from five specialized agent layers: intelligence, research, content, experience, and performance.
• Intelligence agents monitor brand reputation across many sources; research agents build customer segments and synthetic personas from survey data.
• Content agents generate marketing materials across channels; performance agents use algorithms, including multi-armed bandit and reinforcement learning approaches, to optimize decisions.
• The orchestration layer, which coordinates all five agent types together, is the framework's most ambitious and still-developing piece.
• The project is open to contributors interested in agentic AI plus generative AI, verifying truth in AI, or optimization and reinforcement learning specifically.

Who this is for

Madison is a good fit for anyone who wants hands-on experience building agentic AI systems with a real, practical application, marketing, rather than a purely academic exercise. It's an open-source research project under Humanitarians AI, and contributors can visit humanitarians.ai to get involved in building out any of its five agent layers or its orchestration engine.