Multi-Agent Supervisor Architecture: Orchestrating Enterprise AI at Scale

BASF is a German multinational and one of the world’s largest chemical companies, known for its integrated Verbund production network, global scale, and broad portfolio spanning from basic chemicals to advanced agricultural solutions. With its strong foundation in research and development, BASF operates across diverse industries while continuously driving innovation and sustainability.

One of its key operational divisions is BASF Coatings, which specializes in developing, manufacturing, and marketing advanced automotive and industrial coatings, along with decorative paints. As a pioneer in eco-efficient surface technologies, BASF Coatings is also at the forefront of digital transformation, leveraging AI-powered platforms to enhance productivity, innovation, reliability, and design.

In partnership with Databricks, BASF Coatings has implemented a production-ready, governed, and business-impacting multi-agent solution. This approach not only enhances cross-team collaboration but also enables smarter, faster decision-making across critical enterprise functions — setting a benchmark for how advanced analytics and AI can drive tangible business outcomes.

The Challenge: Bring more Modularity, Specialization and Control to Agent Systems

As an organization with over 11,000 employees across more than 70 sites worldwide, managing the growing complexity and improving efficiency of cross-department digitalization is a non-trivial task. More specifically, turning vast, disparate organizational data into actionable insights, and enabling real-time decision-making and productivity has become the key. Solving this problem mattered because efficient digital collaboration and data utilization directly affect market responsiveness, innovation speed, customer satisfaction, and operational reliability. The stakes were particularly high in industries like coatings, where agility and precision are crucial amid rapidly changing customer demands and sustainability pressures.

An agentic system - where autonomous or semi-autonomous AI agents proactively manage business processes and data flows - was the best solution because it could automate coordination and analysis tasks that previously required intensive manual effort. Agent systems could empower organisations like BASF Coatings to:

• Seamlessly integrate AI across domains, automating routine operations in sales, procurement, and supply chain management.
• Provide smart, contextual recommendations and automate decision flows, dramatically reducing bottlenecks and errors.
• Improve user experience by enabling “always-on” chat assistants for support, Q&A, or workflow integration across departments.
• Drive adoption of everyday AI tools company-wide, making complex digital capabilities accessible to business stakeholders and fostering data literacy.

As powerful as an agent could be, as we develop these systems, they might grow more complex over time, making them harder to manage and scale. For example, an agent can have too many tools at its disposal and make poor decisions about which tool to call next, also the context grows too complex for a single agent to keep track of. There is a need for multiple specialization areas in the system (e.g. supervisor, domain orchestration, subject matter expert, etc.)

Another way to view the challenge is through the diversity of data that forms the agent system’s knowledge base. Many of us are already familiar with RAG (Retrieval-Augmented Generation), a technique that combines large language models (LLMs) with real-time data retrieval to improve response accuracy and relevance. However, RAG systems are primarily designed to handle unstructured data - Such as documents, web pages, PDFs, or other forms of free text - rather than structured tables with predefined fields and relationships. When working with structured data, Text-to-SQL is the most common approach for natural language analytics. However, it often relies on pre-defined example SQL queries and lacks built-in mechanisms for data governance and permission control.

The Solution: An End-to-End Multi-Agent Supervisor for Structured and Unstructured Information

To address these challenges, we propose breaking our application into multiple smaller, independent agents and composing them into a multi-agent system. This system will follow a supervisor pattern that coordinates the specialist agents - specifically, Genie agents and function-calling agents - which interact with the Databricks Vector Store Retrieval tool.

AI/BI Genie, one of the most popular features within Databricks, is designed to make structured data such as Delta tables and views directly accessible to business users by leveraging natural language interfaces. It utilizes metadata from Unity Catalog, such as table descriptions, PK/FK relationships, and column names/descriptions. This metadata guides Genie in parsing user questions, constructing accurate SQL, and delivering contextually relevant answers - helping to mitigate errors or hallucinations. In addition, Genie authors can enhance the space by locally editing metadata, defining joins, adding synonyms, and curating BASF-specific instructions. This allows data stewards to actively manage and maintain the quality of their Genie spaces thus contributing directly to the agent system with their invaluable business domain knowledge.

To ease the use of Genie within agent orchestration frameworks, there are frameworks supporting dedicated Python wrappers for building Genie agents (check here for reference). In addition, Databricks product team features example notebooks that walk our users through setting up a multi-agent system using Mosaic AI Agent Framework together with Genie. These examples leverage LangGraph (an open-source agent orchestration library) and demonstrate how to compose workflows where Genie is one specialized agent among several.

An overview of our architecture is as follows. We adopt Databricks’ Mosaic AI framework to simplify the complexities of managing AI agent lifecycles, offering tools and rapid multi-agent coordination prototyping, rigorous evaluation, and effective real-time operational monitoring. Notably, we also integrate the deployed supervisor endpoint with Microsoft Teams for real-time agent execution, and make AI-powered insights readily available to all types of users, including business stakeholders who are less familiar with data platforms - by embedding conversational deployment endpoints directly within the Teams interface. Clear, reusable accelerators exist for provisioning cloud resources (Azure Bot Service, App Service) and connecting endpoints to Teams.

Real Business Impact

While BASF Coatings is developing AI agents that can enhance its business processes, the first landing zone project, Marketmind, focuses on the Sales & Marketing division. The use case enables advanced quantitative and qualitative analysis by consolidating internal Salesforce customer visit reports and market consumption insights with external market trends including S&P 500 news. Some of this data is already processed and available in the form of Delta tables and views, while the rest exists as free-text files and PDF documents, each arriving at different speeds and being updated at varying frequencies. Additionally, the data is managed by different teams and stewards. For example, structured tables are primarily provided by BASF’s central Enterprise Data Lake (EDL) organization, with Sales & Marketing business experts enriching them with domain-specific metadata. In contrast, unstructured data is primarily processed through code-first ETL pipelines developed and maintained by the Coatings Data & AI office team.

Given the complexity of the data landscape, we adopted the multi-agent supervisor architecture for the Marketmind project and used the template notebook as our starting point. We created a Genie space for structured data, enriching it with curated tables, detailed column descriptions, Genie-local join relationships, and value sampling. To improve accuracy, we added SQL examples and clear instructions to guide Genie’s responses, and we performed regular Benchmark tests as new data came in to evaluate its overall performance.

For unstructured data such as Salesforce visit reports and market news, we built vector search indices for each source using embeddings to enable context-aware similarity search. We then created Unity Catalog functions that wrap Mosaic AI Vector Search queries, ensuring enterprise-ready governance, discoverability, and automatic MLflow tracing. Finally, we developed a function tool-calling agent that invokes vector retrieval tools to handle task-specific requests passed along by the supervisor.

Our Marketmind project began its scoping phase in April this year, followed by a 5–6 week proof of concept (PoC). We then moved into the full implementation phase, accompanied by technical upskilling workshops, architecture reviews, and product and feature discussions with the Databricks’ Mosaic AI product team. We conducted a one‑month pilot with 25 key users, and are now in the final refinement stage before go‑live and rolling out to North America by the end of October . Once launched, more than 1,000 sales representatives worldwide will be using Marketmind, with inputs updated frequently.

Marketmind is already changing how BASF Coatings’ sales teams prepare, engage, and follow up with their customers. Instead of hunting for leads through scattered notes and folders, sales representatives receive personalized notifications alongside suggested actions and strategies based on current events in the market. If further information is required, Marketmind offers the option to dig deeper into the underlying data and reports using an easy-to-use chat interface. The screenshot below illustrates this shift. Signals from the market are presented in an actionable, conversational interface inside Microsoft Teams, so Coating’s sales team can shift their focus from “What happened?” to “What should I do next?” without switching tools.

As shown above, sales teams can not only ask ad-hoc questions to the Marketmind chatbot directly in Teams, but also receive proactive adaptive cards with the latest market trends on a weekly basis. Users can explore topics of interest in greater detail by clicking the attached URL, which redirects them to the original data source. To further enhance the agent’s quality, we have also integrated a voting mechanism that allows users to quickly give a thumbs up or down, or provide more detailed written feedback in the bottom field. This feedback is captured in the model inference table and integrated with the existing payload data.

“Marketmind turns our field interactions into timely, AI-driven actions—nudging smart follow-ups, surfacing relevant opportunities, and connecting peers facing similar challenges. The result: faster prep, sharper customer conversations, and more time selling where it counts.” — Adrian Fierro, Head of Global Market Intelligence at BASF Coatings

Why It Worked

Multi-agent architecture with Genie as an agent offers several significant advantages for enterprises like BASF that look to leverage AI effectively in their business contexts. We conclude the key strength into the following aspects:

Specialized agent capabilities with high scalability and modularity: within a multi-agent system, various agents can focus on their specific domains or tasks, enabling deeper expertise in handling diverse queries and datasets. Moreover, organisations like BASF can expand their gateway to AI solutions with an architecture that allows each business division to operate independently while being centrally orchestrated. This modular design helps manage complexity over time.

Enhanced collaboration and improved user experience: agents can share information and context with one another, allowing for more comprehensive responses that integrate data from multiple sources. This facilitates smarter, faster decision-making across various enterprise functions. By integrating AI endpoints to MSFT Teams as a chat interface, we allow users to interact with agents using natural language, making it more accessible to non-technical stakeholders.

Governance and compliance: Protecting personal and customer data is the Commented foundation of Marketmind and remains our highest priority. Every interaction is built on strict compliance with BASF’s data protection standards, leveraging Databricks’ enterprise-grade governance capabilities such as Unity Catalog for fine-grained access control, lineage tracking, and auditability. This ensures that while Marketmind accelerates insights and actions, it does so within a secure, transparent, and fully governed environment.

Close team work between BASF, Databricks and partners: From project start, BASF Coatings, Databricks account and product teams, and partner Accenture proactively engaged in workshops,. which helped align business objectives, technical requirements, and product vision, setting a strong foundation for successful implementation. Right on time, hands-on sessions created rapid feedback loops. Expert guidance was continuously provided by Databricks product team, helping to customize the solution for the complex, evolving needs of BASF and ensuring enterprise-grade quality.

Looking Forward: Multi-Layered Orchestration and Agent Bricks

With the success of the Marketmind multi-agent supervisor solution, the company is now expanding the business impact across broader operations, including Supply Chain, Procurement, Chemetall (Surface Technology subsidiary), and People & Culture. Together with our product team, we are exploring a more scalable multi-layered architecture, where each division operates its own multi-agent supervisor, while a higher-level Coatings-wide orchestrator serves all users. This hierarchical system - a “supervisor of supervisors” - strikes the right balance: it enables division-scoped data and tool access control, preserves flexibility in agent development, and supports a Coatings-wide “Ask Me Anything” capability.

One of our future enhancement goals is the adoption of Agent Bricks, introduced this year at the Data & AI Summit. While our current Mosaic AI–based solution supports multi-agent orchestration, it remains code-first and requires a more hands-on approach with added complexity in deployment and management. Agent Bricks offers a streamlined way to build and optimize domain-specific, high-quality AI agent systems for common use cases, including multi-agent setups. With features such as automatic optimization, cost and quality efficiency, and user-driven feedback mechanisms, it simplifies agent implementation and allows teams to focus on core challenges - data, metrics, and problem-solving. Although we have not yet been able to fully test its capabilities due to limited regional availability, we view Agent Bricks as a visionary direction and plan to enable integration once it becomes accessible, accelerating division-specific multi-agent supervisor development.