Agents

The agent responds to users, works with connected knowledge base materials, calls authorized tools, and can help the user navigate the interface.

This section covers the full agent lifecycle: create, train, test, publish, connect tools, and understand why an answer turned out the way it did.

Main tasks

Creating an agent manually - how to start and save a new agent.
Auto-training - how to create an agent foundation from project materials.
Agent settings - name, instructions, autostart, and AI model.
Agent tools - which capabilities and tools are available to the agent.
Agent Knowledge Base - which materials are connected to the agent.
Knowledge Base Permissions - what the agent may read, edit, create, and delete.
Avito listings - how to add listing descriptions and photos to the Knowledge Base and teach the agent to find the correct listing by its exact ID.
Subscription Page - how to configure the link, appearance, and channel subscription.
Agent testing - how to check responses before launch.
Draft, publication and launch - how to bring the agent to a working state.
Project Operator and Agent Replies - how to prevent team messages from triggering the agent and what happens to a response already being prepared.

Create an agent manually or start auto-training.
Complete the Settings tab.
On the Tools tab, connect knowledge sources, data, and the required capabilities.
Configure the Subscription page tab when it is needed.
Test answers to typical customer questions: on a computer, the chat is to the right of the settings; on a phone, use the Testing tab.
Publish the draft and activate the agent. To have it reply automatically in a particular channel, select that channel in Autostart settings. You can also assign the agent manually to an individual dialog.

If answers differ from what was expected, start with the published agent version: check the instruction, knowledge sources, data and variables, MCP servers, and tool permissions. To check the interface, open the relevant article for the required section and compare the steps with the current screen.