Supervisor
Intelligent Orchestrator
The Supervisor agent emerges as an orchestration mechanism that enables management of multiple agents and/or flows within a single Project, depending on the elected governance type. This agent intelligently analyzes incoming interactions, dynamically selecting the most appropriate agent for each scenario.
For example, a technical support interaction might involve multiple agents: a diagnostic agent identifying the issue, a troubleshooting agent providing solutions, and a billing agent handling any related account queries—all coordinated seamlessly by the Supervisor agent, ensuring consistent user experience.
This approach automates complex workflows by:
Routing requests to specialized agents
Preserving conversation context
Implementing reasoning, with intelligent decision-making
Making guardrails, and content and behavior restrictions central
Resolves ambiguities
Controlled digression for smooth transitions between agents with relevant context transfer.
The Supervisor agent is generated automatically by the system after you complete the Project creation.
How to Configure the Supervisor
The Supervisor agent is created automatically alongside the Project. You can access the Agents page to edit at any time or complement it with any additional configuration.
Role
The role of a Supervisor Agent is to act as an orchestrator of a team of AI Agents. Its main function is to analyze user queries, determine which specialized agent is best suited to execute the task, and delegate it accordingly. It comes filled by default.
Goal
In this version, the Supervisor agent Goal comes filled by default as “Responsible for processing user inputs, analyzing requests, and delegating them to the appropriate agent.”, and cannot be edited.
Instructions
Instructions serve as the detailed blueprint for your Supervisor's orchestration behavior and general guidelines. This field should focus on high-level coordination rules and general behavioral guidelines that apply across all interactions, rather than specific task execution details. It can be a long form text or bullet lists.
This field is crucial for dictating the behavior of the Supervisor as an orchestrator. If not well described, orchestration may be inefficient and fail.
In this field, you can include:
General behavioral guidelines and protocols
How to handle out-of-scope situations
Escalation procedures
What the Supervisor should or should not do during orchestration
Avoid including: Specific task procedures, detailed workflows for individual functions, or step-by-step instructions for specialized tasks.
The Supervisor delegates tasks to specialized agents - it shouldn't contain detailed instructions about how to perform specific tasks (like opening tickets, processing payments, etc.), as these belong in the specialized agents' Goals field.
Guardrails
Present in both Supervisor and Agent configurations, Guardrails provide safety mechanisms to prevent inappropriate outputs or harmful behavior.
Guardrails establish essential safety boundaries that govern what your agents can and cannot do. These protective parameters ensure that all interactions remain appropriate, secure, and aligned with organizational values. Guardrails prevent potentially harmful outputs while maintaining the agent's effectiveness within its designated role. They can control content filtering levels, limit access to sensitive information, and define appropriate response protocols for various scenarios.
In this section, we should only include restrictions on how the agent should handle data, always taking into account its scope. It is recommended to avoid instructions related to other agents or routing commands, as this point is designed to define security aspects, such as instructing the agent not to expose sensitive information.
You can choose to have Agents inherit these preferences from the Supervisor or override them with agent-specific Guardrails.
Constraints
Rules allows to define structured and organized instructions, segmented by topics or sections, serving the same purpose as general instructions but in a more controlled and explicit format. Through this field, you can establish functional constraints that define the agent’s operational boundaries, such as which topics to avoid, which actions are not allowed, and which limitations must be acknowledged, ensuring the agent remains focused on its intended domain and operates within clearly defined capabilities.
Just like Guardrails, Constraints appear in both Supervisor and Agent settings and can be inherited or overridden as needed.
Advanced Settings
Use these hyperparameters to help you to customize the language model behavior. You can set the request timeout, level of creativity combining two or more parameters, and control vocabulary variation and repetitions.
Default Persona
This is where you define the default Persona. Having one for all AI Agents allows you to maintain a consistent style, tone, and language throughout the conversation. If you have more than one in your Project, you can change it here at any time. Just keep in mind that changes to a Persona in use will affect the outputs generated by all the agents using it.
Macroprompt
The macroprompt presented below serves as the framework that guides how the Supervisor analyzes user inputs, makes decisions, and delegates tasks to specialized agents within the project. This setup not only enhances the model’s visibility into applicable rules and constraints, but also standardizes its behavior to ensure coherence, efficiency, and alignment with the defined objectives.
Last updated
Was this helpful?