Generative orchestration updates coming to Microsoft Copilot Studio! 

Update 10/3: We have fixed an issue with new orchestrator capabilities that affected some agents using generative mode. Please try again and we appreciate your patience whilst we addressed this issue.

We’re excited to announce some significant updates to the generative orchestration capabilities for custom agents in Copilot Studio. When enabled in a custom agent, generative mode uses generative AI to select the right actions and topics to respond to your users and can fully generate any conversation needed to gather required information and provide an answer.

During the public preview so far, we have seen enormous excitement about how generative orchestration can transform both the way you build agents, as well as how it can transform the experience for your customers or employees – with your agent able to now handle multiple intents at one time and handle significantly more use cases than was possible with topics and traditional NLU alone.

We have been listening hard to your feedback and have made some key changes to generative mode for agents.

Agents can now use knowledge more seamlessly

It’s already incredibly simple to add knowledge to your agent, giving it the ability to search and summarize answers, when no suitable topic or actions can be found. Now though, with generative mode enabled, knowledge can be used alongside your topics and actions, meaning your agent can choose to your knowledge as the best place to find an answer immediately. Even better, this also means that knowledge can be used in combination with your actions and topics, allowing your agent to use knowledge to populate missing inputs.

Unified, more natural, responses

Until now, when an agent uses one or more actions to respond to a user’s query, each action will generate an appropriate response. Whilst powerful, this still meant that agents produced multiple, individual, responses. We have worked hard to improve this and make conversations even more natural and now your agent will generate a single, unified, response using all the information returned from the actions and knowledge identified.

This change also now allows your topics to be fully generative too. Just add the inputs and outputs via the topic details tab and your agent will generate the necessary conversation to fill the inputs and use your output to generate a response, reducing the need for explicitly authored question nodes and enabling your topic’s responses to be much more flexible.

Improved accuracy and performance

We have made some significant changes to how agents identify appropriate topics, actions and knowledge to use, meaning you should see improved accuracy and significantly reduced latency, vs right out of the box.

Try these changes and give us your feedback

The great news is that these updates are rolling out right now and all you have to do to try them is to enable generative mode in your custom agent’s settings.

Hint: If you’re already using generative mode in your agent, the only change you should consider making is to disable sending a response once your action has been called, since your agent will handle that for you now in most cases.