Glossary
Terms as this documentation uses them. Where a word has a loose industry meaning and a precise Arble meaning, the precise one is given.
| Agent | A reusable configuration: a model, a set of tools, and instructions. Not a running process — that is a run. |
|---|---|
| Run | One execution of an agent. Has a status, a deadline, and usage. |
| Session | A conversation with accumulated context. Runs happen inside sessions. |
| Tool | One typed capability the agent can call, with an input and output schema. |
| Skill | A packaged procedure that calls tools. A way of working, not a capability. |
| Connector | An authenticated link to an external service, contributing tools. |
| MCP | Model Context Protocol. A standard for exposing tools to any compatible client. |
| Tool registry | The single index of every callable tool, regardless of source. |
| Planner | The component that picks which tool to call for a step. |
| Permission gate | The check every tool call passes before executing. |
| Capability | A permission subject, such as filesystem.write. |
| Scope | The narrowing on a capability — a path, a host, a repository. |
| Memory | Durable per-project facts, retrieved semantically. Outlives sessions. |
| Project | The unit that scopes memory, permissions and connectors. |
| Workspace | Several projects sharing memory and connectors. |
| Profile | An endpoint, organization and credential under one name. |
| Worker | The sandboxed process that executes tool code. Holds no database credentials. |
| Resource | A typed non-scalar tool input or output — a file, image, table or database handle. |
| Requires action | A run paused waiting on you, for an approval or a tool result. |
| Drain | Letting in-flight runs finish before a process shuts down. |