> For the complete documentation index, see [llms.txt](https://docs.borndigital.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.borndigital.ai/getting-started/how-agents-work.md).

# How agents work

How a conversation travels through the platform – from the phone or the chat bubble, through the orchestration layer, to the AI models and your systems – and how the new agent runtime and shared conve

This page is about **agents** – Digital Agents and Realtime Agents. Insights runs separately and is described in the [Insights](https://docs.borndigital.ai/insights/introduction.md) section.

Every conversation with your agent passes through three layers:

1. **Channels** – where the customer talks to the agent: a phone call or the chat bubble.
2. **Orchestration layer** – the part of the platform that runs the conversation: it decides what the agent does in each turn and keeps track of the conversation.
3. **AI and your systems** – the language and voice models, your MCP tools and the knowledge base.

You build the agent in Digital Studio. The platform runs it.

## Three kinds of agents

The orchestration layer can run an agent in three ways. All three use the same channels and the same phone numbers.

| Agent                              | You build it as                                                                                                        | Runs on           | Speech on the phone                                            |
| ---------------------------------- | ---------------------------------------------------------------------------------------------------------------------- | ----------------- | -------------------------------------------------------------- |
| **Digital Agent with a flow**      | A [conversation flow](https://docs.borndigital.ai/digital-agent/conversation-flow.md) with nodes, intents and AI nodes | Flow engine       | Speech-to-text and text-to-speech around the flow              |
| **Realtime agent**                 | A [prompt with skills and MCP tools](https://docs.borndigital.ai/realtime-agent/introduction.md)                       | Realtime control  | A speech-to-speech model hears and speaks directly             |
| **Prompt-and-tools agent** *(new)* | A prompt with skills and MCP tools – no flow                                                                           | Responses control | Speech-to-text and text-to-speech, or the GPT-Live voice model |

```
 Phone (your PBX or our numbers)          Chat bubble
            │                                   │
            ▼                                   ▼
   Voice: speech-to-text, text-to-speech  ◄─────┘
            │
            ├──► Flow engine ──────────────► AI models · knowledge base · your APIs
            │       (Digital Agent flows)
            │
            └──► Responses control ────────► AI models · MCP tools
                    (prompt-and-tools agents)

 Phone ──► Realtime control ──────────────► speech-to-speech model · MCP tools
              (realtime and GPT-Live agents)      │
                                                  └──► Responses control (GPT-Live backend)

 State management ── one shared record of the conversation, used by all of the above
```

### Flow engine – Digital Agent flows

This is how most agents run today. On the phone, the platform turns the caller's speech into text, the flow decides the next step – with intents, conditions and AI nodes – and the answer is turned back into speech. In the chat bubble the same flow answers in text. Everything you build in the flow editor runs here, unchanged.

### Realtime control – realtime agents

Realtime control runs [realtime agents](https://docs.borndigital.ai/realtime-agent/introduction.md): a speech-to-speech model listens and answers directly, without a separate speech-to-text and text-to-speech step. It accepts the call, gives the model your prompt, skills and MCP tools, and ends the call when the conversation is over. It also runs **GPT-Live** agents, where a fast voice model talks to the caller and hands the real work to a backend model.

**Where realtime agents are configured:**

* **In Digital Studio:** provider and model (OpenAI, Azure, Gemini), reasoning effort, voice (model voice or ElevenLabs), transcription, listening and turn detection, no-input reminder and hang-up, prompt, skills and MCP servers, phone numbers and **Go live**, the test call (**Start talking**), analytics in Insights, users and history. An agent set up this way can be tested right away, and its calls are logged for analytics like any other.
* **By Born Digital in the realtime management console, for now:** GPT-Live agents, the protected greeting and goodbye, waiting through silence and per-model tool settings for Gemini. Digital Studio support is coming.

### Responses control – prompt-and-tools agents *(new)*

Responses control runs agents that are built from **a prompt, skills and MCP tools instead of a flow**. Each turn, the model reads what the customer said, decides which of your tools to call – for example to look up an order or book an appointment – and answers. On the phone, Voice turns speech into text and back, as for flow agents. Responses control is also the backend model of GPT-Live agents.

* Changing the agent means changing its prompt, skills or tools – no nodes to rewire.
* It uses the latest OpenAI and Azure OpenAI models through their Responses interface.
* A phone number belongs to exactly one agent. Moving a number between a realtime agent and a prompt-and-tools agent takes it away from the old one.
* Born Digital can follow every turn and tool call of a running conversation, which makes problems quick to find.

Set up by Born Digital on request.

### State management – one record of the conversation *(new)*

State management keeps **one shared record of each conversation**, whichever channel and whichever kind of agent served it:

* every call and chat that belongs to the conversation,
* the transcript in the right order,
* facts learned on the way – for example that the customer has already been verified, or a contract number.

It does not run conversations; it remembers them. Facts can expire, and sensitive facts are hidden from history and search.

Today GPT-Live and prompt-and-tools agents use it to resume a conversation. Flow agents, the chat bubble and other channels will write into it in later releases – that is what will let one conversation continue from one channel and one agent to another.

## What this means for you

* **Your flows keep working.** Digital Agent flows run exactly as before. The new services run next to them.
* **A faster way to build voice agents.** A prompt-and-tools agent goes live with a prompt and your MCP tools, without building a flow.
* **Calls can resume.** A GPT-Live or prompt-and-tools call that drops can continue where it stopped.
* **Coming next – one conversation, many channels.** A customer will be able to start on the phone and finish in the chat bubble or on WhatsApp, with the next agent knowing what was already said and done – even when a different kind of agent takes over.

{% hint style="info" %}
**Availability:** responses control and state management run on the **customer-test** environment from release 21.10.2026. They are not yet available in production. Set up by Born Digital on request.
{% endhint %}
