Multi-agent chatbots
Discover multi-agent chatbots at AI Chatbot Hub: AI systems with multiple agents for complex, interactive user experiences.
Last updated
Discover multi-agent chatbots at AI Chatbot Hub: AI systems with multiple agents for complex, interactive user experiences.
Last updated
AI Chatbot Hub stands as the first and most powerful LLM-native conversational AI platform utilizing a multi-agent architecture. This design allows multiple AI Agents to work together in coordinated orchestration, enabling you to develop solutions with unmatched flexibility, speed, and automation.
This guide introduces you to how our Agents collaborate, along with best practices for designing multi-agent chatbots.
AI Chatbot Hub includes two types of Agents:
User-facing: These Agents interact directly with users in a conversational Q&A style. Only one user-facing Agent is active at a time to respond to user queries.
Background: These Agents do not interact with users directly but instead continuously monitor the conversation. All background Agents activate whenever a user submits a new query to the chatbot.
All Agents operate in one of two states:
Active: The Agent is “Connected” and actively running.
Inactive: The Agent is “Disabled” and does not engage.
Since only one user-facing Agent engages with the user at any given time, our system selects the most suitable user-facing Agent based on their unique areas of expertise. This selection is managed by an AI Supervisor, which uses intent classification to identify the best-suited Agent.
Agents function independently, meaning you cannot influence one Agent’s behavior, selection priority, or bias from another.
For our AI Supervisor to accurately determine which Agent is best suited for each query, the “Agent Description” is crucial. Each User-Facing Agent should have a precise, well-defined description.
Effective Agent descriptions should answer the question, “What types of queries should this Agent handle?” Specifically, they should include all relevant “user intents” the Agent is designed to manage.
Agent descriptions should avoid detailing “how the Agent performs its job.” Such specifics belong in the Agent’s base prompt. For instance, phrases like “Agent provides informed answers based on knowledge from the source library” are unnecessary and may distract the AI Supervisor.
A well-designed multi-agent chatbot ensures that each user intent is handled by the most appropriate Agent. For insights on refining Agent intents, refer to our article on Fine-Tuning Agent Intents.
Our AI Supervisor’s routing design balances latency, consistency, and comprehensiveness. While it may not always achieve 100% accuracy, effective design can get you very close.
If the chatbot’s responses aren’t meeting your expectations, it could be due to one of two reasons:
The wrong Agent was assigned to handle the user’s query.
The selected Agent isn’t optimally configured.
To check if the correct Agent was chosen, go to “Inbox” and enable debug mode.
In debug mode, under “Active Agent,” you’ll see the Agent that generated the response. Note that this metadata is visible only if more than one user-facing Agent is active.
If the wrong Agent responded, you can update its name, description, or prompt in the configuration. If the right Agent responded but the answer was not as expected, the issue lies in the Agent itself. You may need to adjust its configuration. One quick fix is to directly revise the AI’s answer manually.
This adjustment ensures that the chatbot will use your revised response whenever it encounters the question again, provided the same Agent responds.
For greater robustness and consistency, consider refining the training data, selecting a more advanced LLM, or enhancing the clarity of your base prompt. For further guidance, refer to our Best Practices for Preparing Training Data.
