Maestro agents are AI agents tailored to specific business needs, use cases, and teams. Data consumers interact with Maestro agents through Maestro Assistant. You build and maintain custom Maestro agents in Maestro Studio.
Types of Maestro agents
There are 2 types of Maestro agents:
- Collibra-managed agents, which are provided and maintained by Collibra and available by default. Maestro Assistant is the primary Collibra-managed agent and includes Collibra-managed subagents. As these agents are read-only, they can search and retrieve information, they can't create or edit assets, assessments, data quality rules, data contracts, workflows, or agents.
- Custom Maestro agents, which are created by your organization using Maestro Studio. You can add action tools to your custom Maestro agents that create or edit content, subject to your approval. Data consumers use custom Maestro agents through Maestro Assistant.
Maestro Assistant
Maestro Assistant is the primary Collibra-managed agent. It powers the conversational interface where data consumers can ask questions of Collibra-managed and custom Maestro agents, and use custom agents to perform actions.
Admins can update the welcome message and example questions in the Maestro Assistant configuration, and choose the Collibra-managed assistant agents that Maestro Assistant uses in Maestro settings.
Data consumers can start a conversation with a custom Maestro agent in the following ways:
- Type the agent's handle, such as @finrep, in a chat message. The handle works similarly to tagging a colleague in a chat window.
- Select the agent from the drop-down list.
A custom Maestro agent must have the Published status before data consumers can interact with it. Agent owners can also converse with the agent while it is in the Pending Review or Draft status.
Custom Maestro agent components
Custom Maestro agents are specific to the environment they are created in and consist of the following basic components:
- Agent identity
- The agent identity refers to the basic information of the Maestro agent. It includes the name, description, and a unique handle.
- The name and description help human users identify an agent and its purpose.
- The handle is a shortcut name for the agent, such as @SalesAI. Data consumers can type the handle directly in a chat to call that specific agent. The handle works similarly to tagging a colleague in a chat window.
- Agent instructions
- The agent instructions define the behavior and tone of the Maestro agent. They give the agent context and are directly injected into the agent's system prompt.
- Agent knowledge base
- The agent knowledge base restricts the Maestro agent to specific views in the organization or applications. It defines the Collibra data available to the agent. You can expand the content the agent has access to by adding tools, such as Search, to the agent.
- Agent tools
- The agent tools enable the agent to perform actions, in addition to accessing information.
- Maestro test
- The agent calling information defines how the Maestro agent is introduced to data consumers. It includes the welcome message, at least one example question, and the color assigned to the agent.
- When data stewards manage an agent, they can test the agent interactions when defining the calling info. Testing an agent is important and ensures the instructions and the knowledge base return the expected result.
- Agent access
- The agent access defines who can interact with the custom Maestro agent.
- Agent status
- The agent status reflects the current state of the custom Maestro agent in its lifecycle. An agent must reach Published status before data consumers can interact with it from Maestro Assistant, except its owners, who can converse with the agent if it is in the Pending Review, Draft or Published status.
Agent scope
The scope of a custom Maestro agent is determined by the combination of its knowledge base and the tools you give it.
- Knowledge base: Defines the Collibra content the agent prioritizes when answering
questions.
Tip Focus the knowledge base on a specific area of data rather than loading everything that might be relevant. A large knowledge base consumes more tokens and can reduce the accuracy of responses.
- Tools: Extend what the agent can access and do beyond the knowledge base content. For example, allow the agent to search across the broader Collibra Platform or create assets.
The knowledge base and tools work together, not independently. How you combine the knowledge base and tools determines what happens when a user asks something outside of the knowledge base. Depending on your use case, you can configure the agent to stay strictly within its knowledge base, or to search more broadly when it cannot find an answer there.
- Restricted agent
- The agent responds only with the content defined in its knowledge base. It does not search outside that scope. A restricted agent is useful when accuracy and control are the priority. For example, a compliance agent that must only reference certified regulatory data.
To build this pattern, add a knowledge base and do not add a search tool. - Focused agent
- The agent prioritizes a specific area of your Collibra Platform, but can search more broadly when it can't find an answer within that scope. A focused agent is useful when you want the agent to stay focused on a certain set of data, but still handle questions that fall outside it.
To build this pattern, add a knowledge base and add the Asset Keyword Search tool.
Maestro agent tools
By default, a custom Maestro agent can access information from its knowledge base to answer questions. With tools, you can extend the capabilities of a custom Maestro agent to go beyond the knowledge base and search the broader data in Collibra or perform actions on behalf of users.
The tools available to an agent depend on its defined knowledge base. Some tools are only available in specific contexts. You can see the full list of available tools when configuring an agent in the Maestro Studio agent builder.
Types of tools
Tools fall into two broad categories:
- Search and retrieval tools
- These tools allow the agent to look up information in Collibra beyond what is defined in the knowledge base. For example, an agent can search across assets, retrieve asset details, or look up data classifications. When a knowledge base is also configured, the agent checks the knowledge base first and uses search tools to find information outside that scope.
- Action tools
- These tools allow the agent to perform tasks in Collibra on behalf of the user, such as creating assets or business terms. Action tools are human-in-the-loop, meaning the user must explicitly approve the action before it is carried out.
Human-in-the-loop tools
Some tools require the user's explicit approval before performing an action. When the agent is ready to take an action, it presents the user with a summary of what it intends to do and waits for approval or rejection before proceeding. This ensures users remain in control of any changes made to data in Collibra.
The approval request appears in Maestro Assistant.
Example: Search tool
A FinREP agent is configured with a knowledge base containing approved regulatory documentation. A user asks the agent: "Are there any Data Sets related to capital requirements?"
Because this question asks about specific assets in Collibra rather than the regulatory content in the knowledge base, the agent cannot answer from the knowledge base alone. With a search tool enabled, the agent can search across assets in Collibra to find relevant Data Sets and return them to the user, without leaving Maestro Assistant.
Example: Build a custom Maestro agent
An organization collects and governs all European financial regulations reporting data in Collibra. To help data consumers find accurate information, you want to build a custom Maestro agent that references only certified data assets in the Finance community and can create missing terms. Maestro Studio enables the creation of this specialized Maestro agent.
- Agent identity
- You configure a new agent with the name "FinREP Assistant", a description, and handle "@finrep".
- Agent instructions
- You define the system instructions indicating that the agent must maintain a professional tone and must always cite regulatory sources.
- Agent knowledge base
- You define the knowledge base by restricting the agent to specific views in the Finance community that contain only FinREP-certified data assets. This ensures the agent can't access or reference non-regulatory data.
- Agent tools
- You select that this agent can use the tools Get Asset Details, Prepare asset for creation, and Create asset.
- Maestro test
- You assign orange as the agent color, add an example question, and add the agent welcome message to explain agent capabilities to users.
- Agent access
- You select that only the Finance group can access this agent.
- Agent status
- You submit this agent for review. Once successfully reviewed, the agent is in the status Published.
The result is a specialized assistant for finance professionals that provides regulatory guidance based on approved and governed data, with the ability to create assets to fill compliance gaps. The agent is available to the professionals through Maestro Assistant. They can start a conversation with the agent by calling the @finrep handle in a chat message or by selecting the FinREP Assistant from the drop-down list.