Glossary
| Term | Meaning | Not to be confused with |
|---|---|---|
| Account | The installation's single trust domain: members, workspaces, conversations, executors, credentials and usage | A tenant; there is no tenant isolation layer |
| Agent | An agent-kind member with a durable identity, stewarded by a human | The program that implements it (the agent application) |
| Agent application | The product runtime implementing an agent's behavior and policy — rho, or yours | Nexus; a runner |
| Approval | The kernel stage every tool call crosses; the mode and rules decide whether it runs, parks or is refused | A permission the runner enforces |
| Awareness | A notification an agent may ignore | Obligation |
| Capture | A file an executor publishes and names in a result | An input attachment |
| Conversation | The durable dialogue and history host | A session; a run |
| Cost unit | The account's configure-once currency for estimates | A price |
| Executor | A delivery address with a transport credential: agent application, runner or tool provider | A member; a physical machine |
| Fork | A new conversation branched from a point in history; a side is a fork mode | Regeneration |
| InferenceRequest | A bounded, sealed direct model call outside a run | A conversation reply |
| Input | Accepted intent awaiting delivery: a message, a steer, kernel mail, a scheduled instruction | A turn |
| Kernel mail | A background result, child reply or scheduled report delivered as a kernel-origin input | The executor inbox; a notification |
| Manager | The human who approved a runner or tool provider | Steward; owner |
| Member | A human or agent account member | The system principal |
| Memory | Model-readable documents scoped to a conversation, workspace, person or answering agent | History; the store; working files |
| Model provider | A configured lane to a model-serving backend | A tool provider |
| Obligation | An inbox task with a claim, a deadline and a write-once result | Awareness |
| Owner | A workspace's one human manager | Account owner (a role) |
| Plane | The one authority axis of a credential: member, executor transport or platform | A scope |
| Profile | An agent's settings and declarations | A second identity |
| Run | One durable execution: a task graph with results | A loop (the algorithm); a turn |
| Runner | An executor serving tools bound to an execution environment | An in-process orchestration class; a sandbox service |
| RunnerEffects | Per-runner evidence of claimed writes: untouched, touched, unavailable | Proof a write completed |
| Schedule | A durable instruction and clock dispatching child conversations | A background job |
| Speaker | Attribution of who spoke on an input or turn | The authenticating principal |
| Steer | An input delivered to the next model boundary of running work | A queued message |
| Steward | The human responsible for an agent | Manager; owner |
| Store | Scoped opaque JSON persistence | Memory |
| Task | One unit of work in a run, delivered to an executor as an inbox row | A second delivery noun |
| Tool provider | An executor serving tools by name from a pool | A model provider; a runner |
| Turn | One slot on a conversation timeline with an opening input and a reply | One model message or provider request |
| Variant | One candidate answer for a turn | A provider attempt |
| Workspace | The required container for conversations, runs and InferenceRequests, with an owner and an access mode | A directory; a machine |