Agent Types

An agent is the configured personality that answers or places a call. Every agent has a voice, an LLM, and a definition of how it decides what to say. There are two kinds.

LLM-driven agents

The common case. The agent is defined by a system prompt plus a set of tools (HTTP Functions and, optionally, MCP tool servers). Each turn, the LLM receives the conversation so far and generates the reply, calling tools as needed. This is the flexible, open-ended mode — good for support, Q&A, and anything conversational.

Flow-driven agents

For structured, deterministic conversations, an agent can instead be a flow — a graph that is walked node-by-node. Each node decides the next based on the caller's input. This trades open-endedness for predictability, which suits scripted flows like verification or menu-driven routing.

How a turn works

When the caller stops speaking, the agent transcribes what was said, decides what to do, and replies in its own voice. The reply starts playing as soon as the first sentence is ready, so the agent does not wait to finish thinking before it speaks.

If the caller talks over the agent, the agent stops and listens. This is called barge-in, and you can tune how sensitive it is in the agent's speech settings.

Tools

More agent features

• Knowledge base: upload .txt or .md files. The agent looks up the relevant passages on each turn. An agent with no documents skips the lookup.
• Post-call analysis: after the call, Dialgem reads the transcript and fills in the variables you define, on the call record.
• Webhooks: signed events for call_started, call_ended, call_analyzed, transcript_updated and the transfer events. See Webhooks.
• Privacy controls: redact personal data from the stored transcript, or turn off transcript storage and call recording.