From One to Many: Multi-Agent Systems
1. From One to Many: Multi-Agent Systems
Welcome back! So far in this course, we've explored numerous single-agent use cases - from AI travel agents planning trips to customer support agents processing refunds. But what happens when workflows grow increasingly complex? Let's explore this with a real-world example.2. Customer Support in the Real World
We’ve seen many customer support examples throughout this course, but in reality, customer support is much more complex than simply answering queries and accessing a few guidelines. Within any customer support team, you encounter many edge cases, use-case categories, and query types. You might have: Billing disputes requiring financial knowledge, technical issues needing engineering support, legal compliance questions demanding regulatory knowledge, product recommendations calling for sales acumen, and more.3. Customer Support in the Real World
Each of these query categories requires access to specific knowledge bases and tools.4. The Problem with Single Agents
Building a single agent that can tackle all these issues can lead to degraded performance. The agent will struggle with increasingly complex logic where models need to be aware of multiple conditions and edge-cases, and tool overload, where agents need to handle a large amount of tools to achieve their objectives.5. Multi-Agent Systems
This is where multi-agent systems come in. Unlike a single-agent system, where a single model equipped with appropriate tools and instructions executes workflows in a loop. Multi-agent systems execute workflows across multiple coordinated agents. Multi-agent systems can be designed in numerous ways for specific workflows, but two broad common patterns have emerged.6. The Manager Pattern
The manager pattern — where there is a central model that orchestrates a network of agents to perform specific tasks. This pattern is ideal for workflows where you only want one agent to control workflow execution and have access to the user.7. The Manager Pattern Applied
In the customer support example, think of this interacting with the head of customer support, who delegates tasks to the appropriate team member, in this case billing, but manages the communication with the client.8. The Decentralized Pattern
The decentralized pattern instead relies on the concept of handoff. Here, a triage agent hands off the request to another agent, who can handle the task end-to-end, including communicating with the user. This pattern is especially effective for scenarios like conversation triage, or whenever you prefer specialized agents to fully take over certain tasks without the original agent needing to remain involved.9. The Decentralized Pattern Applied
In the customer support example, think of this as first interacting with a customer support agent, that connects you with the relevant expert for your use-case, and that expert then owns your ticket.10. Emerging Agent Architectures
There are many emerging architectures for building multi-agent systems. And we’ll cover these in more depth in subsequent courses.11. From One to Many: Multi-Agent Systems
For this course, what you need to know is that as workflows become increasingly complex, you can switch from single agent to multi-agent systems.12. Let's Practice!
Now, let’s practice your skills.Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.