Multi-Agent System (MAS) Design Workshop

Why to use?

Automate a complex process with AI by designing a multi-agent system (MAS) together with an interdisciplinary team in a design thinking workshop.

Who to use

Multi-agent systems require the collaboration of various experts in the following areas:

• Business domain and processes

• Big data and Agentic AI

• IT, security and privacy

It also requires all stakeholders to work together, i.e.:

• Users: employees and/or customers

• Developers: software, data & AI engineers

• Decision-makers: business process owners

• Consultants: privacy officers, security advisors, etc.

Finally, a facilitator is needed to bring the different people together to form an effective team.

When to use?

The team has come together and decided on a (existing or new business) process. The team wants to automate and optimize this process with Agentic AI in order to increase efficiency, effectiveness and/or robustness.

If no process has been identified yet, we recommend a Lean Data & AI Strategy Workshop to identify and prioritize potential use cases for Agentic AI.

During the one-day MAS Design Workshop, participants specify the objectives and key results, identify the human and artificial agents and design the work and information flow between the agents. In addition, the technical and analytical foundations are defined and the necessary guardrails are specified to ensure security, privacy, fairness, and more.

What to use?

This workshop template is based on the tried and tested Data & AI Business Design Method, which is used worldwide by many well-known companies and consultancies. It uses the canvases of the Data & AI Business Design Kit, which is made freely available under a Creative Commons license.

How to use?

This workshop template is designed for a one-day session that can be split into two half-days. Over the course of the day, participants go through the following six phases and perform several steps on six different canvases for each phase. On the canvases, you will find numbers in circles (①, ②, ...) that correspond to the steps in each phase.

I. Intro

The intro is about ensuring that all participants are pursuing the same objective and are willing to follow the proposed path together. To outline and present this path (i.e., workshop agenda), we use the Data & AI Design Thinking Workshop Canvas and carry out the following steps:

① Customize the canvas header by specifying the company and, if applicable, the consultancy, and entering the date. This step should also be performed on all subsequent canvases.

② Set the specific Objective for the workshop and define the desired Key Results (i.e., deliverables).

③-⑦ Adjust the Agenda Items, Session Times, and more as needed.

When you are working on an agenda item, change the color of the corresponding sticky note to yellow. When you are finished, mark it green. This ensures that the team always has an overview of the current status.

II. Business Process Analysis

Next, we need to visualize and analyze the business process that we want to automate and optimize with Agentic AI. To do this, we use the Value Chain Canvas and green sticky notes for existing elements, yellow for planned elements, and red for missing elements.

① In the Focused on field, write down the name of the business process.

② Start at the beginning of the process on the left side of the Value Chain Canvas: a) What is the Initial State, the Base Products, or the trigger that starts the business process? b) Which person, role, or organizational unit defines the initial state, delivers the base products, or triggers the event (Producers)?

③ Then continue with the end of the business process on the right side of the Value Chain Canvas: a) What is the Final State, the End Products, or the key results of the business process? b) Who are the beneficiaries of the final state, the consumers of the end products, or the recipients of the key results (Customers)?

④ Now outline the Primary Activities, i.e., the workflow of the process: What actions are necessary and in what order do they occur? What alternative or parallel flows are there?

⑤ In addition to the Primary Activities, Support Activities are often necessary: which general and internal measures/organizational units support the business process?

⑥ If the support activities are not provided internally but by external companies and are involved throughout the entire process, note these under General Suppliers.

⑦ Special suppliers that only perform or support individual process steps should be placed under Special Suppliers.

⑧ Specify the Key Performance Indicators (KPI) on blue sticky notes, which measure the efficiency, effectiveness, quality, stability, etc. of the process and specify corresponding target values.

Finally, extract all objectives, key results, and KPIs relevant to Agentic AI from the Value Chain Canvas and specify them in the green box Business Objectives & Key Results as the agents' output.

III. User Role Identification

Now that we know and understand the business process, we can answer the question: who are the stakeholders in the context of the process? To answer this question, we use the Stakeholder Analysis Canvas and blue sticky notes to identify the persons or roles. We should consider the following functions of stakeholders (whereby one person/role can also take on several functions):

① In the Focused on field, write down the name of the business process.

② Decision Makers: Who makes decisions that influence the flow of the process?

③ Economic Buyers: If purchasing, budget, or other (financial) resource decisions are made during the process, who provides the money or is the sponsor?

④ End Users: Who uses the results of the process (cf. Customers on the Value Chain Canvas)?

⑤ Saboteurs: Who could try to disrupt the process by negatively influencing decision makers, economic buyers, or end users?

⑥ Influencers: Who could support the process by positively influencing decision makers, economic buyers, or end users?

⑦ Recommenders: Who is actively involved in the process in an advisory function or provides support with information?

⑧ Implementors: Who actually implements the process, i.e., performs the actions?

In the later multi-agent system design, agents take on the function of some persons/roles. But even with a very high degree of automation, certain functions remain with human agents: they have to check the results of the AI agents, grant approvals, or are the users and/or beneficiaries of the automated process.

By changing the color of the sticky notes, we mark certain stakeholders:

• Green are human agents who play a role in the Agentic AI process.

• Red are stakeholders who are no longer supposed to play a role (i.e., no human agents).

• Yellow are those stakeholders we are not yet sure about (i.e., maybe human agents).

Finally, we transfer all human agents to the yellow box Human Agents in order to design the Agentic AI workflow and AI agents in the next step.

IV. Agentic Workflow Design

For the agentic workflow design, we use the Diagram Format and the Agentic Workflows Shapes of Miro. We already know the outputs that we expect from the agents from II. Business Process Analysis. From III. User Role Identification we know the stakeholders, who give inputs to the agents, i.e. trigger the process, write prompts, provide information and documents, answer queries from the agents, make decisions or check and approve (interim) results.

There are two options for automating the existing process with AI:

1. Largely retaining the existing process flow and replacing the human agents with AI agents that perform the actions and make decisions.

2. Completely rethink the process flow in order to exploit the advantages of agentic AI: for example, the parallel processing of several alternative solutions.

If you are not sure which variant is better, design two (or more) versions and then decide - or mix the solutions.

To identify candidates for AI agents, ask the workshop participants the following questions:

• What activities or responsibilities have the human agents taken on?

➡️ The human agent becomes an AI agent.

• What specialized tasks are there for which special domain knowledge is required?

➡️ An AI agent is trained with this domain knowledge.

• Which IT systems or data sources do we need to connect?

➡️ An AI agent serves as an interface to the data source or IT system.

• Which user (roles) do we need to interact with?

➡️ An AI agent handles communication with the user.

• Which activities can be parallelized?

➡️ The activities are distributed to different AI agents.

• Which activities are required multiple times by other agents?

➡️ An AI agent makes this activity available to other agents as a service.

• Which activities require special security and data protection precautions?

➡️ Specially secured AI agents perform these activities.

• Which internal AI agents are already in use?

➡️ The existing AI agent is reused.

• Which external AI agents are already in place?

➡️ The external AI agent is integrated and, if necessary, encapsulated by an internal AI agent.

• Which AI agents can support the coordination of the other AI agents?

➡️ Special AI agents take care of delegation, aggregation, synchronization, etc. of the information and work flow.

To complete the MAS design, the information and work flow between the AI and human agents needs to be modeled. To do this, the diagram elements (human agents, AI agents, deliverables) are connected with arrows. As a rule, the information and work flows are identical. In cases where this is not the case, a dashed line can be used for the pure information flow.

V. Data & AI Assessment and Roadmapping

Some of the AI agents need read access to existing data sources or even write access to IT systems in order to change or create data records or trigger certain sub-processes. Other AI agents require specific functionalities and capabilities such as a Large Language Model (LLM) to communicate with human agents or a predictive model to forecast events and trends.

We use the blue area of the MAS design diagram to specify the required IT / BI / AI systems and connect them to the AI agents using arrows. Here we can also define interfaces such as the Agent2Agent (A2A) or Model Context Protocol (MCP).

Next, we evaluate whether we already have the required systems in use, whether they are still in the planning or working stages, or whether they first need to be designed and developed. We use green, yellow and red sticky notes and the Analytics & AI Maturity Canvas to do this. The canvas differentiates between general tools and specific applications at different levels of complexity and maturity.

② The green boxes describe the specific applications for:

a) Business Operations: pure data processing applications without analytics or AI functionality.

b) Business Reporting: typically applications for automating report generation or dashboards based on descriptive analytics.

c) Business Discovery: applications for exploring trends, correlations, anomalies, etc. to gain insights based on diagnostic analytics.

d) Business Forecasting: Applications for forecasting, nowcasting or backcasting based on predictive analytics.

e) Business Optimization: Applications for the optimization of business processes based on prescriptive analytics.

f) Business Automation: Applications for the automation of business processes based on autonomous analytics.

AI agents often interact with existing applications via application programming interfaces (API).

③ The yellow boxes are intended for the data & analysis tools that can be used to implement the AI agents:

a) Data Management: This includes, for example, database systems.

b) Descriptive Analytics: For example reporting or dashboarding tools.

c) Diagnostic Analytics: Statistical analysis tools or, for example, platforms for A/B experiments.

d) Predictive Analytics: In addition to software for machine & deep learning, also libraries e.g. for Bayesian networks, linear regression etc..

e) Prescriptive Analytics: Methods for simulation and optimization are used here as well as generative AI solutions.

f) Autonomous Analytics: Reinforcement learning algorithms are used here, among other things, or special solutions for agentic AI.

Ensure that there is an existing or planned IT / BI / AI system for all required data and functionalities.

VI. AI Guardrail Requirements

We have worked our way through the business process, the stakeholders, data sources, and IT / BI / AI capabilities to ensure that our multi-agent system is viable, desired, and feasible. Another important criterion is still missing: AI systems must be responsible.

After all, with great power comes great responsibility. This principle is also enshrined in law, for example in the EU AI Act.

To ensure that our AI agents comply with the principles of Responsible AI (rAI), we need so-called AI guardrails. In a multi-agent system, these guardrails can in turn be implemented by agents that monitor and control the other agents.

First, we define the guardrails, i.e., the rules and restrictions we want to impose on the MAS. We use the 3 Boxes Canvas to divide the rules into three categories:

1. Security & Safety: Neither the MAS nor its users may suffer any damage.

2. Explainability & Transparency: Users must be able to understand the MAS's decisions and actions.

3. Privacy & Fairness: Users must not suffer any disadvantage as a result of the MAS.

We also distinguish between guardrails relating to the input, the internal models, and the output of the AI agents:

• Input Rail: Could, for example, check user input for prompt injections to protect company data from data theft.

• Model Rail: One example of ensuring fairness is monitoring model quality indicators to rule out discrimination against groups of people.

• Output Rail: Since LLMs hallucinate, a plausibility check of the output is useful, for example.

In the final step, the Guardrail Agents (grey box on the diagram) define how to implement these rules and how they are connected by arrows to the AI Agents.

To conclude the workshop, specify concrete jobs to be done and assign them to the workshop participants. And then: Get the jobs done!

Where to find more?

Join the free Data & AI Business Design Community to get access to tutorials, exercises and more.

---

Copyright: All rights reserved by Datentreiber GmbH. For more workshop templates, transformation tools, and canvas tutorials, visit our Data & AI Business Design Bench. The Data & AI Business Design Bench requires a commercial licence per user per year. For more information, please contact Georg Arens via Email or make an appointment via Calendly.