AI Agents
Give a bot a real personality and let it reason instead of following a script
An AI Agent is a conversational AI you configure rather than script card by card. Give it a system prompt, a personality, and (optionally) documents to learn from and tools it can call, and it decides how to respond on its own, turn by turn.
You can use an Agent two ways: as a single Agent card inside a Flow (for one step of a larger scripted conversation), or as the thing that runs the whole conversation.
Creating an agent
A step-by-step wizard walks you through:
- Basics: a welcome message, a short list of icebreaker prompts, an avatar, and a color theme.
- Prompt: the system prompt that defines who the agent is and what it should do.
- Tools: optional actions the agent can call mid-conversation (like looking up an order through an API Connection), each of which can require a person's approval before it runs.
Communication Style
A dedicated tab lets you shape how the agent talks without rewriting its whole prompt:
- Answer length (concise / detailed)
- Answer as (paragraphs / bullet points / numbered list)
- Tone (friendly / professional / casual / empathetic)
- Language (respond in a fixed language regardless of what the visitor writes in)
- Formatting (plain text or markdown)
Every one of these defaults to "Let Agent Decide", so a new agent behaves reasonably out of the box, and nothing changes for an agent you've already set up until you deliberately pick a value.
Training the agent
The Training tab is how you give an agent real knowledge to answer from:
- Website: point it at a URL and it learns from that page.
- File Upload: upload PDFs or Word documents.
- FAQ's: write question-and-answer pairs directly, no file needed.
Whatever the source, it's chunked and indexed the same way, so the agent can draw on all of it together when answering.
Preview
The Test tab holds a real, multi-turn conversation with the agent exactly as a visitor would experience it, so you can judge how it behaves across several exchanges, not just on one message. Use Reset to start a fresh conversation without reloading the page.
Advance settings
- Max tokens: a ceiling on how long a single reply can run, to bound latency and cost.
- Fallback model: a second model/provider to retry with if the primary call fails or is cut off at its token limit, so a visitor gets a usable reply instead of an error.
- Search Mode: trades retrieval depth for speed when the agent looks up relevant training content: Fast, Balanced, or Thorough.
Un-answered Questions
This tab shows questions the agent had ready knowledge for and still couldn't find anything relevant to answer, a real signal of a gap in its training, not a guess. It stays empty for an agent with no documents yet, since that's a different, expected situation (nothing to search yet, not a failure to find something).
Use it to spot what to add to Training next.
Meta Data
A description and free-text tags, purely for keeping a growing list of agents organized. They have no effect on how the agent behaves.