Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
These frequently asked questions (FAQ) describe the AI impact of generative orchestration for custom agents built in Copilot Studio.
What is generative orchestration?
Generative orchestration lets your custom agent answer user queries with relevant topics, actions, other agents, and knowledge sources, and respond to event triggers.
Generative orchestration enables more natural conversations by filling in inputs using details from the conversation history. For example, if a user asks about the nearest store in Kirkland, and then asks for the weather there, orchestration infers they mean the weather in Kirkland.
Triggered agents can use generative orchestration to determine the best action, topic, or agent to invoke in response to outside events, enabling autonomous capabilities. For example, an agent can check for and reconcile duplicate accounts when a Dataverse table for sales accounts receives a new entry.
The system can also chain together multiple capabilities—answering queries like "I need to get store hours and find my nearest store"—and ask follow-up questions if any required details are missing or ambiguous.
What can generative orchestration do?
Generative orchestration builds a plan to address a user query or event trigger using the name, description, inputs, and outputs of available topics, actions, agents, and knowledge.
For conversations, the system references the last 10 turns of conversation history to fill in inputs and determine the most relevant capabilities to invoke. It follows up with the user for any missing or unclear details, executes the selected plan, and then generates a response based on the output, including any custom agent instructions.
For event triggers, the orchestration uses event data, trigger-level instructions, and agent instructions to decide which topic, action, or agent to invoke and how to respond.
What are the intended uses of generative orchestration?
You can use generative orchestration to create agents that respond to user queries and events by reasoning over topics, actions, other agents, and knowledge, using the available context and metadata. An agent can delegate parts of a task to other agents that are better suited to handle a particular ___domain or function, enabling modular and scalable designs.
How is generative orchestration evaluated? What metrics are used to measure performance?
We evaluate generative orchestration across the end-to-end process: how well it identifies a suitable plan and executes it to resolve a query or respond to a trigger. Quality assessment by human reviewers covers different prompts, inputs, and configurations.
We assess the system's ability to select appropriate actions, topics, agents, or knowledge sources, how accurately it interprets user intent, and how effectively it filters out malicious or harmful content from users or makers.
What are the limitations of generative orchestration? How can users minimize their impact?
For best results, make sure topics, actions, knowledge sources, and agents have high-quality names and descriptions. Learn about writing effective metadata.
Agents invoked through orchestration—whether internal or external—must be configured correctly and capable of handling the queries or events passed to them. If the receiving agent isn't designed to process a particular task, it might return incomplete or irrelevant responses.
Currently, agents with event triggers use only the maker's credentials for authentication. Actions called by an agent in response to a trigger must also use the maker's credentials. For more information, see data protection for agents with triggers.
What operational factors and settings allow for effective and responsible use of generative orchestration?
Generative orchestration is currently supported in English only. You can test its performance using the test panel, directly in Copilot Studio. You can also add custom instructions to shape how the system selects and uses topics, actions, other agents, or knowledge.
When delegating tasks to other agents, it's important to test the interaction flows to ensure that context is passed clearly and that handoffs behave as expected.
What are actions and how do agents with generative orchestration use them?
Actions allow an agent to perform specific operations or retrieve data to answer user queries or handle events. Your organization, Microsoft, and other partners can create actions. You can configure which actions are available, and customize their metadata to support generative orchestration.
What data can Copilot Studio provide to actions? What permissions do Copilot Studio actions have?
When an agent calls an action, the system sends the required input values. This information can include elements of the conversation history or data from event triggers. When the system orchestrates across agents, it supports continuity by passing relevant context to the receiving agent.
What types of issues might arise when an agent uses actions and other agents?
Errors might occur when the agent prepares inputs or generates outputs, or if it selects the wrong action, topic, or agent. To prevent such issues, ensure that metadata is accurate and unambiguous across all elements available to orchestration.
Information from triggers or user queries might include unintended or sensitive data. If such information is routed to a topic, action, or another agent, it might lead to undesired outputs. For more information, see Troubleshooting and limitations.
What protections does Copilot Studio have in place for responsible AI?
Copilot Studio includes a range of safeguards:
- Agents only use knowledge, actions, topics, and agents explicitly configured by the maker.
- Admins can restrict which actions and agents are available.
- Makers can require user confirmation before executing actions that modify data.
- Triggers and orchestration operate under the maker's authentication and are subject to configured permissions.
- Payload inspection, classifiers, and content filters detect malicious or harmful instructions in user input, trigger data, action outputs, and knowledge content.
- Entity validation with Power Fx expressions can constrain input values (for example, limit email recipients to a specific ___domain).
- You can configure if the full conversation history is passed, and the task to complete, when your agent delegates to other external agents.
If a potential attack is detected (for example, in trigger payloads or action outputs), execution is blocked and a content filtered error appears on the activity map.
To maintain transparency, agents include the default message: "Just so you are aware, I sometimes use AI to answer your questions."
How can I give feedback on generative orchestration?
You can provide feedback for Copilot Studio.