Core concepts
Agents
An agent is the thing that talks. It has instructions, a model and a set of capabilities. Each agent is created as a voice or a chat agent, which decides the settings the dashboard shows for it: a chat agent needs no voice and no phone number, a voice agent adds those.
An agent comes in one of two shapes, chosen when you create it:
| Prompt agent | Workflow | |
|---|---|---|
| Defined by | One prompt | A graph of steps |
| Best for | One job, conversational | Branching, scripted flows |
| Control | The model decides | You decide, step by step |
A prompt agent is one prompt doing one job. Fastest to build; the model decides how the conversation goes. A workflow is a graph of steps you connect yourself, for conversations that must follow a path - qualify, then branch, then hand off. See Prompt agents and Workflows.
The dashboard offers the workflow builder for voice agents only; a new chat agent is a prompt agent. The API accepts a chat agent in workflow mode.
Sessions
A session is one conversation - one phone call, one web call, one chat. Every session is stored with:
- the full transcript, turn by turn
- a configuration snapshot taken when the session started: the prompt, first speaker, model, voice and transcriber it began with (for a workflow, those of the first conversation node in its graph; for a realtime agent, its realtime settings)
- what it cost, split by component
- latency, so you can see where the time went
- the recording, if recording is on
When an agent gets something wrong, the session is where you start. See Sessions and transcripts.
Versions
An agent always runs its current version, which is the one you last saved. There is no separate deploy step: an edit is live as soon as it is saved.
Publishing marks the current version as published, records who did it, and sends
a workflow.published webhook. The first edit after a
publish opens a new draft version, so published versions are kept as history;
further saves before the next publish change that same draft.
Publishing does not gate what runs. If an agent is taking live traffic, every save reaches callers straight away. To work on a change without affecting them, duplicate the agent and edit the copy.
A new session starts on the current version, unless you open it through the sessions
API with a specific version to test an older one. After that:
- a call (phone or web) keeps the version it started with until it ends.
- a chat session keeps the same version record too, but if that record is the draft you are editing, a save changes it in place and the next turn of the chat uses the edit.
The Playground
The Playground is where you test an agent before real callers meet it. It can:
- chat or talk to an agent live in the browser
- compare two agents, or two versions of one agent, side by side on the same input. Comparison is text only; a voice call runs one agent at a time.
- replay a real past conversation's messages against a version, to check a change did not break what already worked
- try overrides - a different prompt or model - without saving them to the agent
- seed variables for the session, as your application would
Providers
Glytos does not lock you to one vendor. The language model, the transcriber and the voice are each chosen per agent from a catalog, and you can bring your own keys for any of them, in which case that provider's cost is yours and Glytos does not charge for it. See Models and voices.
Tools and knowledge
Two different ways to make an agent useful beyond talking:
- Tools let it do things - call your API, look up an order, trigger a workflow in another system - during the conversation. See Tools.
- A knowledge base lets it answer from your own documents. See Knowledge base.
Channels
An agent runs on any of:
- Phone - inbound and outbound, on a number you connect from your carrier. See Phone numbers.
- Web - a call from a browser, over WebRTC. See Web calls.
- Chat - the agent as text, in threads. See Chat.
The prompt, model, tools and knowledge base work the same way everywhere. Some settings only mean something on a call (voice, transcriber, silence handling, voicemail) and have no effect in chat. A few capabilities depend on the channel: a client tool runs only in a browser web call, because it needs the browser on the other end; on the phone or in chat it returns an error to the model instead.
Organizations
An organization is a workspace: its own agents, numbers, credit and members. One account can own several and switch between them. Each organization is created in a region, which is where its data lives and cannot be changed afterwards. See Organizations and regions.