AI agents

Explore the model of what an Agentalk agent is, how its components work together, and how calls are processed in real time.

An Agentalk agent is a voice AI system that can make and receive phone calls. To use it, you need to understand how its components work together during a live conversation. This guide explains the six core components of an agent and how the voice pipeline processes a call.

The six components of an agent

Every agent is built from the same six components:

Component	What it does
Instructions	Define the agent's goal, personality, and communication style.
LLM	Generates responses based on the caller's input and instructions.
STT	Converts the caller's speech into text.
TTS	Converts the agent's response into spoken audio.
Functions	Trigger actions during the call based on conditions.
Knowledge base	Provides information the agent can use to answer questions.

Instructions, LLM, STT, and TTS are required for every agent. Functions and a knowledge base are optional and depend on your use case.

How components work together

All six components are active during a call. Audio is converted to text, processed, and converted back to speech in a continuous loop.

How the voice pipeline works

During a call, the agent processes audio and text through a sequence of steps. This process is known as the voice pipeline.

Agentalk supports two pipeline approaches, each with different trade-offs in latency, flexibility, and cost.

Chain model

The chain model connects three separate systems in sequence:

1. The STT model listens to the caller and converts speech into text.
2. The LLM reads the transcript, applies the agent's instructions, and generates a response.
3. The TTS model converts that response into audio and plays it back to the caller.

When to use the chain model

Use this model when you need flexibility. You can choose different providers for each stage, optimize costs, or prioritize specific language support.

Real-time model

The real-time model uses a single system to handle listening, reasoning, and speaking together.

There is no handoff between components, which reduces latency and makes conversations feel more natural. This model typically costs more than the chain model.

Cost vs experience tradeoff

Real-time models provide better conversation flow but are more expensive. Choose based on your priorities.

Temperature

Temperature controls how closely the agent follows its instructions versus how much variation it introduces in responses.

• Low temperature: The agent stays consistent and follows instructions closely. Use this for compliance or structured scripts.
• High temperature: The agent responds more flexibly and varies its wording. Use this for more natural conversations.

Default recommendation

Start with a moderate temperature and adjust based on how consistent or flexible you want the agent to be.

Inbound vs. outbound agents

An agent behaves differently depending on whether it is making or receiving a call.

• Outbound agents initiate the call. They should introduce themselves immediately and clearly state the reason for calling.
• Inbound agents receive calls. They should greet the caller and first understand the caller’s intent before guiding the conversation.

Use different scripts for inbound and outbound calls

Inbound and outbound calls require different opening messages. Using the same script for both results in a poor caller experience.

Create separate agents for inbound and outbound scenarios. You can duplicate an existing agent to speed up setup.

Functions

Functions are actions the agent can trigger during a call based on what the caller says. You define the trigger condition in plain language, and the agent decides when to activate it.

There are four function types:

Function	What it does
End Call	Ends the call when the caller indicates the conversation is complete.
Update Contact	Extracts information from the call and stores it in the contact record.
Transfer Call	Forwards the call to another phone number.
External API	Sends or retrieves data from an external system during the call.

How functions are triggered

You don't call functions manually. The agent decides when to trigger them based on the conditions you define.

See Agent functions for details on how each type works.

Knowledge base

A knowledge base lets the agent answer questions using your own content.

During a call, the agent searches the knowledge base for relevant content and uses it to generate a response.

When to use a knowledge base

Use it when your agent needs to answer detailed questions from large or complex content such as documentation, product data, or internal knowledge.

When a knowledge base is not needed

Simple workflows such as lead qualification or appointment booking usually don't require a knowledge base.

See Set up a knowledge base for setup instructions, and Connect a knowledge base to an agent to link one to your agent.

Next steps

Here are the recommended next steps after completing this guide:

• Create an agent: Build your working agent and test a call.
• Write effective agent instructions: Define behavior, tone, and conversation flow.
• Use built-in functions: Add actions like call transfer and contact updates.
• Set up a knowledge base: Enable your agent to answer questions using your content.
• Credits and pricing: Understand how usage is calculated and configure your plan.
• Setup checklist: See the full recommended configuration sequence.