Core concepts
| Term | Meaning | Example |
|---|---|---|
| Agent | A task executor with a model, instructions and tools | Support assistant |
| Profile | Named model and authentication settings | default, deep-analysis |
| Workflow | Mattermost routing settings and agent instructions | Assistant for a support channel |
| Session | Conversation context and working environment | Handling one ticket |
| Run | One assignment within a session | Prepare the initial reply |
| Clarification | New information for a running agent | Additional error details |
| Sandbox | Environment containing files and tools | Codex in AgentBox |
| Template | Prepared sandbox software | Codex, Python and a helpdesk CLI |
| Hook | Preparation or finalization script | Publish a reply file |
| Namespace | Integration or process label | helpdesk |
| Orpheus | Service that creates sessions and executes runs | Orpheus API and worker |
| Space | Separate service and web interface for recurring tasks | Morning reports |
| Space schedule | Prompt, repeat rule and agent settings | Report at 10:00 on weekdays |
| Occurrence | A due schedule instance with dispatch and run information | Monday’s report |
A session allows one run by default. For conversations, set allow_multiple_runs: true at creation so subsequent runs retain context. Runs execute sequentially, and clarifications during execution work in either mode. Configuration is captured when a session is created. Editing a profile does not reconfigure an existing conversation.
Reusable sandbox context remains available until the paused sandbox is deleted after its retention period. Session history remains readable after deletion.