Supervisor-Style, The king of Multi-Agent Systems

Author(s): ravindu somawansa

Originally published on Towards AI.

Supervisor-Style, The king of Multi-Agent Systems

This year is shaping up to become the year of Agents 😍. There is not a single day without the arrival of a new Agent based tool or a Agentic framework. At the same time, Agent-based tools are also making their way to the public with tools like OpenAI’s or Perplexity’s Deep Research.

But did you know that, in the most complex Agent-based tools, a specific type is used ? The Supervisor-Style Multi-Agent (yes, it is quite a mouthful 😅).

So in this post, let’s deep dive into Supervisor-Style Multi-Agent systems and how powerful they are.

What is a Multi-Agent System?

Before explaining what is a Multi-Agent System, we need to define first an Agent.

An agent is an entity capable of autonomously achieve a given task. Agents, as we talk about, are based on the ability of LLMs to take decision given some context. The bigger or more specialised the model is, the “better” the decision.

A multi-agent system is simply a team of agents that collaborate to achieve autonomously a goal. It needs to be specific to a use case:

• Specialised agent: Each agent is specialised for a particular task. The more precise this task is, the better it will performs. For example, you will have a specialised agent to search the web and get relevant information or a specialised agent to write the output in the given format.
• Specific Tools: Each agent, depending on its task, should have access on the needed tools and only those. The more tools you give, the more errors it can do.
• Specific model: Each agent should have a specific model behind it, tailored to the task. The more complex the task is, the bigger the model we need but will cost more.

Different types of Multi-Agent System

There are different types of topology for Multi-Agent System, from the Single Agent (which is what everyone use in chat like ChatGPT or Gemini) to the most complex Hierarchical System where you will have multiple layers of agents to build enterprise scale systems.

One of the features that differentiate Multi-Agent systems is how the hands off is done. The hands off is how the next agent is chosen when one has finished its tasks. There are 3 different types:

• Custom hands off: This is what happens in the Custom Agent System where the hands off can be materialised by a directional graph. Each agent can only hands off to specified ones depending on a condition. For example, in Agentic RAG, if the librarian Agent does not find any data in the knowledge base, then the web search Agent will search for relevant information. If there are data, then the main agent will answer the question using those.
• Swarm hands off: This is what happens in Network Agent System. When an Agent finishes its task, it chooses which Agent comes next. This is really powerful when the agents have really specialised roles and the collaboration is simple.
• Supervisor hands off: This is what happens in all the other Agent systems. Every time a specialised agent finishes its task, it returns its results to the Supervisor Agent that will decide of the next steps.

Now that we understood Multi-Agent systems, let’s deep dive on the Supervisor-style!

What is a Supervisor Style Multi-Agent System?

Supervisor Style Multi-Agent System are well suited for complex use cases where you need a plan to achieve a given goal. Most of the more complex Agent Systems are Supervisor-style because of the following features:

• Planning : At the beginning, the Supervisor generates a plan to accomplish a given goal. The more better its domain knowledge is, the better the plan will be.
• Re-Plan: At each agent turn, the supervisor can update the plan depending on the returned information by the other Agents.
• Backtracking: If the current plan direction is deemed wrong, the Supervisor can decide to backtrack and go back to a previous state and try a different direction.
• Human Feedback: At the beginning and at each agent turn, the Supervisor can ask human feedback. It can ask feedback to refine the user’s goal, choose a specific direction or have more context.

This is all possible because you have a Supervisor Agent that plays the role of the Project Manager who handles the communication with the angry clients and coordinates its desperate teammates.

Strategy vs Tactics with Supervisor

The classic Supervisor style is what is used in chat systems like ChatGPT or Gemini, you have the Supervisor (which is also the central Agent) that decides which other agents or tools to use and everything passes through it. This system is as powerful as the capacity of the Supervisor to plan and use its agents and tools.

But by itself, it is limited because the Supervisor will always decide on the most straightforward usage of agents to achieve a tasks. It will not think about using advanced steps like reflection to analyse what has been generated, ranking of different hypothesis or novelty checking for a new hypothesis. You need to explicitly write it to the Supervisor. At this point, you are telling how the Supervisor should, you give it a workflow.

A workflow is just a high level description on how to achieve a goal. Giving a workflow to a Supervisor is just giving it the overall method to complete a task but you let it decide on the specific steps. You give it a strategy but let it decide the tactics.

One of the best example is Google’s AI Co-Scientist which aims to collaborate with researchers to accelerate scientific discovery (from months of work to days or hours 🤩). It use a Supervisor-style Multi-Agent System where collaboration between agents are done in stage where only specific agents work together. This stages are actually modeled on human interaction which is one of the reasons why this kind of Multi-Agent Systems are so powerful.

👉 Want more information on Google’s AI Co-Scientist ? Then check out my post on this awesome Multi-Agent System.

Why is the Supervisor-Style with Workflow so Powerful ?

The Supervisor-style with workflow is this powerful because it is modeled on how a human team works:

• Role based Specialised Agent: The Agents are not specialised for a specific task but for a specific role inside a team. It is modeled on how a real team would work on a task.
• Test Time Compute: One way to have more complex answer from an LLM is to make it “think longer” before answering. This is called Test Time Compute Scaling and it is one of the basis for Reasoning Models. By using this principle, we can make each Agent “think” more deeply on what it needs to do and have “better” and more novel answers.
• Collaboration by Stage: The Agents are not working in a reactive fashion (like in Swarm-style) but in a specified manner. The whole process is divided in Stage where specific roles play their part. Iterations between Stages and inside them happen until the output is acceptable, exactly like a workflow with humans.
• Supervisor as Project Manager: The supervisor plays the role of the Project Manager that interface with the users and coordinate its teammates. It will decide on the iterations to do, when to stop the process, when to ask feedbacks from the users, just like a real Project Manager.

Integrating workflow with Supervisor-style Multi-Agent System allow us to create systems that mimic on real human teams work. This makes the system more robust and more debuggable because we know that this way works. At the same time, this accelerates the process by a dozen times because it is done by computers.

Conclusion

Supervisor-style Multi-Agent systems (still a mouthful) is a really powerful Agentic architecture for complex use cases. By using the Supervisor Agent as the project manager and modeling the other specialised agents on human teams, these systems can handle the most complex goals while still being incredibly faster than their human counterparts. Now that we know that such systems work, the next challenge will be to tailor them to the different use cases of each company. 😤

👉 If you enjoyed this article and want to read more about AI, reasoning models, and Multi-Agent systems, follow me here on Medium or connect with me directly on LinkedIn!

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI