Configure real-time agents

Note

Features in this article are powered by the standard harness, which uses the billing options described in Licensing for agents powered by the standard harness. Learn how to access standard features in Access standard agents and agent flows.

To deploy a real-time agent, enable the real-time model, configure the core agent settings, and then set up the required channel-specific features. You can deploy the same agent to the voice channel, digital messaging channels, or both.

Note

Real-time agents is in preview for digital messaging channels. Preview features aren't meant for production use and might have restricted functionality. These features are subject to supplemental terms of use. Microsoft provides them before an official release so that customers can get early access and provide feedback.

Set up and enable real-time agents

1. Create a new agent, and configure its basic details, such as a descriptive name and the purpose of the agent in the description.

2. Go to the agent's Voice settings and turn on Enable voice. In Voice type, select Real-time.

   Important

   This setting is a one-time selection. After you select Real-time, you can't switch back to a basic agent type. To use a basic agent, create a new agent. This setting is required even if you're using the agent on a digital messaging channel.

3. Select the model you want to use, GPT-Realtime, GPT-Realtime-Mini, or GPT-5-Chat (Preview). GPT-5-Chat (Preview) is a text LLM voice model that generates voice responses through Microsoft Neural Text-to-Speech (TTS).
   Learn more about how these models work based on their Supported regions and deployment considerations. The following table describes the scenarios in which you can use different models.

   Scenario	GPT-Realtime	GPT-5-Chat (Preview)
   Conversation style	Native speech-to-speech interactions	Speech is converted to text for reasoning and then synthesized back to speech
   Latency	Lowest latency for natural conversations	Higher latency than GPT-Realtime
   Voice customization	Supports the built-in real-time voice options available for real-time voice agents	Supports Microsoft Neural Text-to-Speech (TTS), including a broader voice catalog and custom voice models
   Branded voice experiences	Limited customization options	Recommended when you need a custom or branded voice experience
   Regional deployment flexibility	Limited to supported real-time model regions	Greater flexibility through speech recognition and TTS services
   Best suited for	Natural, low-latency voice conversations and open-ended interactions	Scenarios requiring advanced voice customization, custom voices, compliance requirements, or regional deployment flexibility

   Tip

   • Use GPT-Realtime when natural conversation quality and low latency are the primary requirements.
   • Use GPT-Realtime-Mini (Preview) for scenarios that need fast speech interactions with lower latency and resource usage.
   • Use GPT-5-Chat (Preview) when voice customization, custom voice models, or deployment flexibility are more important than response latency.

4. Go to the agent's Security settings and select No Authentication.

Knowledge and tools

You can configure your agent to use knowledge and tools. Learn more in Knowledge sources summary, Add tools to custom agents, and Tools, knowledge, MCP, and API.

Nested agents (preview)

Real-time agents only support child agents.

Important

Ensure child agent descriptions don't overlap with topic descriptions. Explicitly define invocation order in the agent's instructions.

Topics

Real-time agents support all topics configured in Copilot Studio. Use topics to define deterministic behaviors such as greetings, business rules, and escalation, while the model you selected manages conversational responses at runtime. Learn more in Choose how to control the conversation.

Best practices

• Use topics only when deterministic behavior is required.

• Use static text in greeting messages for the fastest first response. Dynamic messages with variables and expressions increase initial latency.

• Disable the Conversation Start topic if you want the real-time agent model to handle the greeting. Otherwise, the greeting configured in the Conversation Start topic plays instead of the voice model greeting.

• Let the real-time agent model handle general conversation and follow-up questions.

• Include an explicit action in the On Error topic, such as transfer or end call. Message-only error handling isn't sufficient. Without a deterministic next step, customers might experience silence for voice agents or stuck calls for messaging channels, leading to confusion and poor customer experience.

• Use explicit topic and tool descriptions to declare ownership of data collection. Learn more in Write effective topic and tool descriptions.

Topic node support

The following list describes topic support in real-time agents:

Condition node

Feature	Support
If/Else branching	Supported
Power Fx expressions	Supported
Slot filling reprocessing	Supported

Message node

Feature	Support
Basic message	Supported
Message variations	Supported
Variable insertion	Supported
SSML	Voice channel only
Rich Media/Adaptive Cards	Supported only for digital messaging channels. This feature is in preview.
Quick Replies	Supported only for digital messaging channels. This feature is in preview.

Question node

Feature	Support
Prompt text	Supported
Automatic hold	Not supported
Slot filling	Supported
Skip behavior/Greedy slot filling	Supported
Reprompt/Retry	Supported
Invalid response handling	Supported
Topic interruption	Supported
Barge‑in	Supported for voice channel only.
Custom reprompt message	Supported
DTMF input	Supported for voice channel only.
Silence detection	Supported for voice channel only.

HTTP node

Feature	Support
HTTP methods: GET, POST, PUT, PATCH, DELETE	Supported
URL endpoints	Supported
Headers and payloads	Supported
Response parsing and schema	Supported
Variable mapping	Supported
Error handling	Supported

Tool node

Feature	Support
Power Automate flow	Supported
Tool invocation	Supported
Input/Output mapping	Supported
New prompt	Supported

Set variable value node

Feature	Support
Literal assignment	Supported
Expression assignment	Supported
Variable to variable	Supported

Topic management node

Feature	Support
End current topic	Supported
End all topics	Supported
End conversation	Supported
Go to step	Supported
User input for recognize intent	Supported
Go to another topic	Supported

Transfer conversation node

Feature	Support
Transfer to agent	Supported
External phone number transfer	Supported for voice channel only.

Advanced

Feature	Support
Create generative answers	Supported

System trigger support

Trigger	Support	Details
On Conversation Start	Supported	Fires when a new conversation begins
On Talk to representative	Supported	Transfers to human agent
Unknown Topic/On Unknown Intent	Not supported	Fallback when no topic matches
OnSelectIntent (multiple topics matched)	Not supported	Disambiguation between similar topics
Reset Conversation (OnSystemRedirect)	Supported	Clears variables and restarts flow
On Sign in	Not supported
Unknown DTMF key press	Supported	Unmapped keypad input. Applicable to voice channel only.
The agent chooses / User says a phrase	Supported	Agent selects topic based on intent
A message is received	Not supported	Increases latency
A custom client event occurs	Not supported	Only at session start
The conversation update	Not supported	Members added or removed, session changes
It's invoked	Not supported	Requires synchronous UI
It's redirected	Supported
The user is inactive for a specified period.	Supported	User inactive timeout. Applicable to digital messaging channels only.
Silence detection	Supported	Applicable to voice channel only.
A plan completes	Not supported
AI response generated	Not supported
On Error	Supported	Handles orchestration errors

Pass variables between topics and the language model

When you use topics in a hybrid conversational flow, understanding how to pass variables between topics and the real-time language model is critical for building reliable, stateful interactions. This process applies across channels.

This functionality works through the following process:

• You pass input variables defined on a topic into the topic at invocation time, so the language model can provide structured data to the deterministic flow.

• You return output variables defined on a topic to the language model at the end of topic execution as structured key-value pairs.

• Tool call outputs follow the same pattern.

• The language model is populated with conversational context, including tool call output key-value pairs. However, you only return explicitly defined output variables as structured data.

• The variable description helps the model understand how to use the variable during a conversation. Provide clear and descriptive definitions.

Learn more in Manage topic inputs and outputs.

Multilingual support

Add all secondary languages you want. Localization strings aren't required for any real-time flows. However, for deterministic topic messages, you need to provide the translated messages. Learn more in Configure and create multilingual agents.

The real-time model can understand and respond in many languages. However, Microsoft doesn't formally validate all languages for general availability.

As of June 2026, the following languages are formally validated:

• Dutch (Netherlands) (nl-NL)
• English (Australia) (en-AU)
• English (United States) (en-US)
• English (United Kingdom) (en-GB)
• French (Canada) (fr-CA)
• French (France) (fr-FR)
• German (Germany) (de-DE)
• Italian (Italy) (it-IT)
• Portuguese (Brazil) (pt-BR)
• Spanish (Spain) (es-ES)
• Spanish (United States) (es-US)

Microsoft continues to validate other languages and adds them after certification completion. You can add any language supported by Copilot Studio. However, languages that aren't fully certified for GA-level quality should be thoroughly tested before production deployment.

Important

Technical language capability doesn't equal a supported or certified language. If you intend to deploy agents in languages other than English, you should conduct extensive testing with real-world callers and call flows before going live.

Context variables

A real-time agent supports context variables that allow it to behave more intelligently by carrying information about the conversation and the customer. The available context variables depend on the channel. This set includes:

Context variable	Description
Channel ID	Identifies the communication channel used for the interaction. Applicable to voice channel only.
Caller phone number (ANI)	The originating phone number of the caller. Applicable to voice channel only.
Callee number (DNIS)	The destination phone number that the caller dialed. Applicable to voice channel only.
Conversation ID	A unique identifier for the active call session.
SIP headers	Supported SIP header key-value pairs associated with the call. Applicable to voice channel only.
Current date (UTC)	The current date in Coordinated Universal Time (UTC), provided at runtime to allow date-aware responses.
Current time (UTC)	The current time in Coordinated Universal Time (UTC), provided at runtime to allow time-aware responses.

Learn more about all other context variables, including digital messaging and custom context variables, in Configure context variables for agents.

Channel specific settings

Voice agent settings

Agent voice

Select the voice your agent uses by selecting your agent and go to Settings > Voice > Select voice. Real-time agents support the following voices:

• Alloy
• Ash
• Ballad
• Coral
• Echo
• Sage
• Shimmer
• Verse
• Marin
• Cedar

Note

• The agent voice is for your real-time agent and isn't the one configured in Copilot Service admin center.
• To match your Dynamics system message voices with your real-time agent, use only the following supported voices: Alloy, Echo, Shimmer, or Ash.
• Agents that use Text LLM, GPT-5-Chat (Preview), generate voice responses through Microsoft Neural Text-to-Speech (TTS). This model supports a broader voice catalog and custom voice models, making it suitable for organizations that require branded voice experiences or advanced voice customization.

Speech sensitivity

Speech sensitivity voice activity detection (VAD) determines when the agent should respond after the caller finishes speaking.

Understanding VAD types

Real-time agents support two VAD approaches:

Server-based VAD - Based on sound (silence)

• Detects end of speech based on audio signals (silence duration, volume)

• Responds quickly once silence is detected

• Best for structured interactions, short responses, noisy environments

Semantic VAD - Based on sentence context

• Determines turn completion based on meaning of what was said

• Adapts to natural pauses, filler words, trailing speech

• Best for: Conversational interactions, complex questions, open-ended discussions

Select the right VAD

Use server-based VAD when all of the following conditions are true:

• Interactions are structured (IVR-style menu navigation).
• Responses are short and predictable.
• Background noise is a concern (semantic VAD might wait too long).
• You want fast, crisp turn-taking.

Use semantic VAD when all of the following conditions are true:

• Conversations are open-ended.
• Callers might hesitate or use filler words ("um", "let me think...").
• Questions are complex (callers explain situations).
• Natural conversation flow is prioritized.

Configure server-based VAD

Go to Settings > Voice > Phone Setup > Speech input > Sensitivity > Based on sound (silence).

Parameter	Description	Default	Recommended range
Threshold	Sensitivity to voice versus noise (0-1 scale)	0.6	0.5-0.7
Prefix padding (ms)	Audio captured before speech starts	300 ms	200-500 ms
Silence Duration (ms)	Silence required to end turn	750 ms	750-1000 ms

Threshold

• Lower (0.3-0.4): More sensitive; picks up quiet speech, might trigger on background noise.
• Higher (0.7-0.9): Less sensitive; requires louder speech, reduces false triggers.
• Recommended: Start with 0.5; increase if background noise causes false triggers.

Prefix padding

• Captures audio before speech detection (prevents cutting off first word).
• Lower (200 ms): Faster response; might miss first syllable.
• Higher (500 ms): Safer capture; slight delay.
• Recommended: 300 ms (good balance).

Silence Duration

• How long the caller must be silent before agent responds.
• Lower (500 ms): Fast turn-taking; might interrupt if caller pauses mid-thought.
• Higher (1000 ms): More patient; might feel slow.
• Recommended: Start with 750 ms.

Configure Semantic VAD

Go to Settings > Voice > Phone Setup > Speech input > Sensitivity > Based on sentence context.

Parameter: Eagerness (how quickly the agent responds after semantic completion)

Setting	Behavior	Best for
Low	Waits longer, very patient	Callers who think out loud, frequent pauses
Medium	Balanced (default)	General conversations
High	Responds quickly	Fast-paced interactions, simple questions

DTMF configuration

Dual-Tone Multi-Frequency (DTMF) allows callers to enter information by using their phone keypad.

You can turn on DTMF for your agent at both the topic and global levels. To set it at the global level, select your agent and go to Settings > Voice > Conversation behavior > DTMF.

Important

When building DTMF-only experiences, configure DTMF input for every input node, including menu selections and multi-digit inputs such as account numbers or PINs. Ensure all possible customer input paths have corresponding DTMF mappings so users can navigate and complete the conversation using keypad input alone.

Silence detection

Silence detection enables real-time agents to recognize when a caller provides no input for a specified period. Set up silence detection as a global voice setting for the agent by going to Settings > Voice > Conversation Behavior > Silence detection.

Important

• Silence detection isn't turned on by default. If the user doesn't speak, the agent waits indefinitely without prompting. Explicitly turn on silence detection and configure a reprompt message to handle silent callers.
• The default silence detection timeout is 7,000 ms (7 seconds). Validate this value against your specific use case and caller environment before deploying to production. Seven seconds might feel too long for some callers or too short for others depending on the nature of the interaction, for example, complex questions or noisy environments. Test with real-world call data to determine the appropriate threshold for your scenario.
• Before enabling silence detection, ensure that the behavior you configure in your silence detection topic (for example, Escalate, Hang Up, Reprompt) is intentional and appropriate for your use case. Misconfigured fallback behavior, such as inadvertently setting the fallback to Escalate when the intent is to hang up, or vice versa, can result in unexpected call outcomes.

Latency messaging

Add latency message or music to your agent when background operations take longer than expected. To configure latency messaging, go to Settings > Voice > Conversation behavior > Latency messaging.

Important

Because the Text LLM real-time agent solution uses multiple sequential steps (ASR, LLM, TTS), callers experience a few seconds of silence between speaking and hearing a response. Set appropriate expectations with users, for example, by using a phrase such as "One moment please, I'm looking that up for you...".

Real-time agent evaluation

Real-time agents support sending text during evaluation, however, audio processing isn't supported.

Digital messaging agent (preview) settings

These settings apply when you deploy the agent to a Digital Messaging channel.

Messaging behavior

Real-time agents on Digital Messaging channels communicate through text. The conversation uses the same AI model, but voice-specific features such as DTMF, VAD, barge-in, and voice selection don't apply.

Inactivity handling

Set how the agent responds when a customer stops responding in a messaging conversation. Use inactivity triggers to define the behavior of your agent when the user is inactive.