Skip to main content

AI agents

An agent combines instructions, knowledge sources, skills, and tools to perform a specific function. The Agents page lets you search, filter, start conversations and, with write permissions, create or edit agents.

Exploring agents​

The tabs separate common and specialized agents. You can search by name or description and filter by creator.

Each card shows its state and available actions: start a chat, edit, activate or deactivate, create a notice, or delete. The organization's default agent cannot be deactivated (All your documents).

Creating or editing an agent​

Select Add agent, or open the three-dot menu in the top-right corner of a card and select Edit agent. The form organizes configuration into four sections.

1. Information​

  • Name: required and recognizable to users.
  • Description: explains its purpose and supports up to 150 characters.

2. Knowledge​

An agent can combine several sources, including sources of different types:

  • Documents: choose a source and limit it to folders or documents.
  • SQL: choose a connection and the tables it can query. If no tables are selected, all are available.
  • API: select a configured API connection.
  • MCP: select a server and its authorized tools.
  • Integrations: add connectors available to the agent.

Select only what is required. A narrower source helps produce relevant answers and reduces unnecessary access.

3. Instructions and skills​

Write the agent's goal, limits, and response format in Markdown. The field supports up to 32,768 characters. GESCON can suggest a draft when you select Generate. It builds a prompt from the selected sources; review it before saving to ensure it matches the agent's objectives.

You can edit this configuration as plain text and render it as Markdown by selecting Preview.

This section also lets you assign skills from the library or upload compatible skill files.

4. Configuration​

  • Groups: limits which teams can use the agent.
  • Maximum turns: between 3 and 50; the default is 15. Increase it only when the agent needs more steps to complete its tasks.
  • Features: allow, require, or disable web search, deep research, document processing, expert mode, and file upload.

A required feature is enabled for every conversation with the agent; a disabled feature cannot be enabled from chat.

Saving and testing​

Review the summary, save the agent, and start a conversation from its card. Test representative questions and confirm that it:

  • Uses the correct sources and cites relevant content.
  • Rejects out-of-scope tasks or indicates that it cannot access information that was not included.
  • Follows the requested format and level of detail.
  • Shows only the features that should be available.

Permissions​

  • Agent Configuration Read allows users to view agent configurations and the agent list.
  • Agent Configuration Write allows users to create and edit agents.

Final visibility also depends on assigned groups and the sources each user can access.