Multi-agent design patterns

Who this is for and what you will get

This is for developers who already run several agents, or who are weighing whether to split one agent into several. After reading, you should be able to judge whether a task needs multiple agents, list the agents it needs, decide how they communicate and coordinate, and see how they interact.^[1]

When multi-agents make sense

The multi-agent design pattern lets multiple agents work together toward a common goal. It is used in robotics, autonomous systems, and distributed computing.^[1]

Three kinds of work fit it well:^[1]

Large workloads. Split the work into smaller tasks, hand them to different agents, and process them in parallel. Large data processing is an example.
Complex tasks. As with large workloads, break the task into subtasks and let each agent specialize in one aspect. Autonomous vehicles are an example: separate agents manage navigation, obstacle detection, and communication with other vehicles.^[1]
Diverse expertise. In healthcare, separate agents can handle diagnostics, treatment plans, and patient monitoring.^[1]

What multi-agents gain over a single agent

A single agent is often enough for simple tasks. Once tasks get complex, multiple agents offer three advantages:^[1]

Specialization. Each agent focuses on one kind of task. An agent that does everything can get confused by a complex task and end up doing work it is poorly suited for.
Scalability. Adding agents is easier than making one agent carry more and more.
Fault tolerance. If one agent fails, the others keep working, so the system stays reliable.^[1]

Take a trip booking. A single agent would need tools for finding flights, booking hotels, and renting cars, which makes the system large and hard to maintain or scale. Separate agents for each job make the system more modular and easier to maintain.^[1]

Six building blocks to settle before you design

A multi-agent system rests on six building blocks. For each one, the question to answer is below, with the trip example.^[1]

Communication. Decide which agents share information and how. The flight agent and the hotel agent must share travel dates so the hotel matches the flight.^[1]
Coordination. Make sure user preferences and constraints are met. If the user wants a hotel near the airport and rental cars are only available at the airport, the hotel agent and the rental car agent must coordinate.^[1]
Agent architecture. Decide how each agent makes decisions and learns from interactions. A flight agent might use a machine learning model to recommend flights based on past preferences.^[1]
Visibility. See the section below.
Architecture pattern. Choose among centralized, decentralized, and hybrid designs, whichever best fits the use case.^[1]
Human in the loop. Tell the agents when to ask a person, such as when a user wants a hotel or flight the agents did not recommend, or wants to confirm before a booking is made.^[1]

Seeing what the agents are doing

Visibility into multi-agent interaction is essential for debugging, optimizing, and checking that the system works. Three tools help:^[1]

Logging and monitoring. Log every action an agent takes. A log entry can record the agent, the action, the time, and the outcome, which supports debugging and optimization.^[1]
Visualization. A graph of information flow between agents makes interactions easier to read and helps you spot bottlenecks and inefficiencies.^[1]
Performance metrics. Track how long a task takes, how many tasks finish per unit of time, and how accurate the agents' recommendations are.^[1]

For the trip example, a dashboard could show each agent's status, the user's preferences and constraints, the interactions between agents, and the flights, hotels, and rental cars each agent has recommended. At a glance, you can tell whether the agents are working together and whether the user's requirements are being met.^[1]

Three common patterns

Group chat. Each agent represents a user in a group chat and exchanges messages through a messaging protocol. The design can be centralized, with all messages routed through a central server, or decentralized, with messages passed directly. Typical uses are team collaboration, customer support, and social networking.^[1]
Hand-off. Each agent represents a task or a step in a workflow and passes work to other agents according to predefined rules. Typical uses are customer support, task management, and workflow automation.^[1]
Collaborative filtering. Agents with different expertise work together to make recommendations. In the stock example, an industry expert, a technical analysis agent, and a fundamental analysis agent each contribute, and together they give the user a fuller recommendation.^[1]

Practice: split a refund process

A refund process can be split into two groups of agents. The first group is specific to refunds: a customer agent starts the refund, a seller agent processes it, a payment agent returns the money, a resolution agent handles problems that come up, and a compliance agent checks the process against regulations and policies.^[1]

The second group is general-purpose and can serve other parts of the business. A shipping agent, for instance, can return a product to the seller during a refund and also handle ordinary shipping after a purchase. Feedback, escalation, notification, analytics, audit, reporting, knowledge, security, and quality agents also fall in this group.^[1]

The exercise is to design a multi-agent customer support system. Identify the agents, their roles and responsibilities, and how they interact, and include both process-specific and general agents. You may need more agents than you expect. Start by thinking through the stages of customer support, and the agents any system needs.^[1]

Limits and common questions

Not every task needs several agents. For simple work, one agent is enough, and splitting too finely adds coordination cost.^[1]
The more agents you have, the more the communication and coordination design matters, and visibility has to keep up, or problems are hard to trace.^[1]
A person in the loop is almost always needed. The key is to spell out when the agents must ask, rather than leaving that judgment to them.^[1]