Help

FAQ: agents

The questions people ask about agents before they know which page to open, from what an agent is and what it costs to who can edit it and how long its conversations are kept.

What is an agent, and does it cost anything while idle?

An agent is one saved configuration in your workspace: its persona, its model, its voice, the knowledge it may read, the tools it may call and the channels it answers on. An idle agent costs nothing; a conversation is charged for its connected time on a call and for its AI usage, and every conversation uses the settings saved on the agent at that moment. A live agent does count against your plan's active-agent limit until you archive it. See What is an agent.

How do I create an agent?

On the agents list press Build New Agent and pick one of three routes: Create with AI interviews you in six questions and designs the agent, Start with a template starts from a written persona for a common job, and Start from scratch asks only for a name and a voice. Every field a route sets is editable afterwards on the agent's detail page. See Create an agent, Build an agent with AI and Quickstart: your first agent.

Which name does the caller hear?

The name written inside the Personality text on the Persona tab. The Name field is the fallback the agent uses when the prompt names no one, and Description is for your colleagues only; the caller never hears it. Write the name into the prompt, as in "You are Sara, the support agent for Acme". See Write the persona.

Can I give each channel its own prompt?

No. One agent has one Personality, and every channel it is connected to gets it; a Zendesk, Zoho Desk, Intercom or Genesys Cloud chat conversation adds a short channel note about tone and the handoff signal, and everything else is your text as written. Run a second agent when a channel needs different wording. See How the prompt is assembled.

How do I set the language the agent replies in?

The voice fixes the accent and the prompt fixes the words, and they are set in two places. Pick a voice on the Voice tab for the accent, since each voice in the library carries one dialect, and write the language rule into Personality, including what to do when the caller switches language. To serve several dialects, add a voice per dialect and turn on Dialect detection. See Voices and language and How the prompt is assembled.

Why does my agent ignore a document I attached?

Retrieval starts off on a new agent, so turn on Enable retrieval on the agent's Knowledge tab first. If it is on, check in order that the source or its folder is selected for this agent, that its status reads Ready, that Relevance threshold is not filtering it out (the default 0.7 is deliberately strict), and that the file has real extractable text, which the Chunks column shows. See The agent did not use my document.

Should reference material go in the prompt or in knowledge?

In knowledge, for anything longer than a couple of lines or anything that changes. The prompt is read on every turn of every conversation and every character of it is charged; knowledge is searched during the conversation and only matching passages reach the model. Keep Personality for rules that are true on every turn. See Add text directly and How the prompt is assembled.

Which kind of tool do I need?

  • One endpoint of your own with a handful of parameters: a webhook tool.
  • A whole API with many operations: an OpenAPI tool, imported once from its document.
  • Answers from your own data: connect a database and approve the questions it may answer.
  • An action inside a product you already connected, such as a calendar event: an integration tool from that connection.
  • Looking something up online, adapting to the caller's dialect or ending the conversation: a built-in tool, already on every agent and switched off until you turn it on.

Tools overview compares the kinds.

Why does my tool never fire?

For a webhook, integration or built-in tool the description is the only thing the agent has to go on, so rewrite it to name the situations you want the tool used in and what comes back. An imported OpenAPI operation has to win a search on the caller's words first: open the tool from the agent's Tools tab, press Test Tool with a sentence a caller would say, and fix the operation's description if it scores below the threshold. See Tools overview and The retrieval bench.

How does my endpoint know a call came from EchoNova?

Only from the credential you attach. Requests to a webhook tool or an imported API carry no EchoNova signature and nothing that identifies the conversation, so treat the endpoint as publicly reachable: require an auth connection or a secret header parameter, and reject every call that does not carry it. See Webhook tools.

Where do I keep an API key or password a tool needs?

Under Settings in the agents navigation, never in Personality or a tool description. A workspace secret is one value that becomes a header you name, for an API key; an auth connection is a whole credential with its method, for basic, bearer or OAuth2, and always produces the Authorization header itself. The secret half of either is not shown again after you save it, so name it for the system and the environment it belongs to.

Can the agent write to my database?

No. Every approved question is a read, the agent never composes SQL, and it can only fill the parameters of the questions you approved. Give it a dedicated read-only database user with SELECT on the tables it needs and nothing else, because the connection string cannot be read back or edited once saved. See Database tool security.

Does testing in the playground cost credits?

Yes, at the same rates as a customer conversation, per turn, with no test mode. A Call test costs more than a Chat test of the same length because a call also pays 2,500 credits for every minute it is connected. Every playground session is recorded on the agent's Conversations tab and counted in its analytics like any other conversation. See Test your agent.

What am I charged for in a conversation?

A call pays 2,500 credits for every minute it is connected, silence included, which covers listening and speaking. Every turn, on a call or in a chat, also pays for its AI usage: the tokens the model read and wrote at the provider's price plus 10%, turned into credits. A tool call adds a model round trip rather than a price of its own. See What a turn costs.

Does changing a setting affect a conversation already running?

No. A conversation finishes on the settings it started with, and a saved change to the persona, the tools or the knowledge reaches the agent at the start of its next conversation; retrieval settings can take up to a minute. Save, then start a new conversation from Test to hear it. See Persona, When a change reaches an agent and How knowledge works.

Where do I read what the agent said and what it cost?

On the agent's Conversations tab. Open a row to read the transcript and the inspector beside it, which shows Total credits, Tools triggered and the summary the agent wrote at the end. The search box matches the customer name and the contact id, not the transcript text. See Find and read a conversation and Transcripts and summaries.

Does pausing an agent stop its channels?

No. Pause Agent archives the agent, which stops the sessions you start from the portal, disables editing and frees the agent's place on your plan; a channel bound to an archived agent still answers messages and calls, and those turns are charged. To stop a channel, disconnect it or assign it to another agent, then pause the agent. See Agent status.

Do working hours stop the agent answering outside the window?

No. Working hours on the Settings tab record when you intend an agent to be available; they do not stop an agent answering outside the window today, and those conversations are charged. To have the agent say it is closed, put the hours and the out-of-hours behaviour in Personality; to stop it answering, disconnect its channels or bind them to another agent. See Working hours.

How long is data kept, and can I delete a conversation?

Nothing an agent makes expires on a timer, and there is no call audio: a voice conversation keeps the seconds transcribed and the words. Conversations are kept until you delete them: open one and press Delete. Agents have no delete; archive the agent instead. See Data retention.