Use Deep Reasoning models for complex tasks.

[This article is pre-release material and is subject to change]

Your agents can use Deep Reasoning to better understand and complete complex tasks. This capability enhances their decision-making abilities and leads to more accurate responses.

You can add Deep Reasoning models to agents that have orchestration and Deep Reasoning enabled.

Note :

Adding Deep Reasoning models to Copilot Studio is currently only supported in the United States and the European Union (excluding the United Kingdom). Both regions use the Azure OpenAI o3 model.

The Deep Reasoning models in Copilot Studio are in preview and there are no commitments regarding where data is stored. Your data may be transferred to, stored, and processed in the United States or any other country where Microsoft or its subcontractors operate.

For organizations using a pay-as-you-use payment model: While Deep Reasoning is only available in certain regions, processing pay-as-you-use Copilot Credits can occur outside the agent's region.

After you enable Deep Reasoning for your agent, the agent will automatically determine which tasks require Deep Reasoning, or you can add the keyword "reason" to the instructions to have the agent apply Deep Reasoning to specific activities.

Important note :

This article contains preview documentation for Microsoft Copilot Studio and is subject to change.

Preview features are not intended for use in a production environment and may have limited functionality. These features are provided before the official release so you can access them early and provide feedback.

How does Deep Reasoning work?

Deep Reasoning uses advanced AI models to complete tasks requiring logical reasoning, problem-solving, and step-by-step analysis. This model allows your agent to complete complex tasks with higher accuracy, albeit slower response times, compared to the default models used by Copilot Studio.

Your agent will determine which tasks or steps would benefit from applying the Deep Reasoning model. You can also instruct your agent to use Deep Reasoning to complete tasks by using the keyword "reason" in the agent's instructions.

Note : Using Deep Reasoning will consume your paid Copilot Credits.

Enable access to Deep Reasoning models

To use inference with an agent, you must enable orchestration generation and grant that agent access to Deep Reasoning models:

1. Open the settings for your agent.
2. Turn on Deep Reasoning (preview) .

Use the Deep Reasoning model to complete the task.

Agents that have access to Deep Reasoning models do not use them for all the tasks or steps they perform. Instead, agents may use Deep Reasoning models to accomplish specific tasks.

Your agent identifies which tasks or steps would benefit from a Deep Reasoning model.

You can also explicitly instruct the agent to use Deep Reasoning for specific tasks. Ask the agent to use reasoning within the agent instruction to complete a specific task or step.

For example, these instructions include the keyword " reason" in one of the steps. The agent only uses the Azure OpenAI o3 model for that step:

1. Analyze the request details to determine: Location (e.g., Arizona) and the services required (e.g., cleaning, gardening, maintenance).
2. Use the "Get suppliers from ERP" process to retrieve internal supplier data.
3. Search your internal knowledge base for information on the supplier's capabilities, location, and performance.
4. Run a web process search and extract Market Share, Revenue, Ranking, and other relevant information from all vendors.
5. Using data and reasoning , provide a top recommendation for the user's request along with relevant context.

Note : An agent can use Deep Reasoning models for multiple tasks or steps, but each task or step must contain the keyword `reason` . Because Deep Reasoning models take longer to respond, using the model for multiple steps can significantly slow down the agent's response time.

Consider the behavior of the inference model.

You can track when and how your agents use Deep Reasoning patterns in the Activity page.

The activity map shows a Deep Reasoning node where the agent used the Deep Reasoning model.

To see the details, select the reasoning node to expand. The expanded node will explain the reasoning steps, the data the model used, and the results the model produced.