Concepts

Agent execution

1. Send the request and conversation to the configured model.
2. Receive text or a tool request.
3. Check the tool request against permissions.
4. Execute an allowed tool and return its result to the model.
5. Continue until the model completes the turn or a limit is reached.

Completed messages and tool results are saved in a session.

Settings and state

Concept	Controls
Operating mode	Interactive, headless, single-reply or server execution
Provider	Endpoint, model and credentials
Tool	File, shell, search and extension operations
Permission	Allow, ask or deny an operation
Profile	Tool visibility, instructions and permission rules
Output style	Reply length and explanation
Session and project	Conversation history and project scope
Memory	Notes available across sessions

See Use cases for task examples and the glossary for definitions.