Glossary¶

The words the console uses, what each one means, and the name the same thing has in the API, the spec and the YAML an agent exports. The console, the onboarding tour, the Architect and these pages use these names; the API and the spec keep their own, so a script written against them never breaks over a rename.

What you build¶

In the console	What it is	In the API and spec
Agent	An assistant with its own instructions, model, knowledge and tools	`agent`
Draft	An agent's unpublished changes; nothing outside the Builder sees them	`draft_spec`
Publish	Freeze the draft as a numbered version that every surface runs	`POST /agents/{id}/publish`
Version	One published state of an agent, kept and comparable	`agent_version`
Test panel	The chat beside the Builder that answers as the draft or any version	`is_test` on its runs
Instructions	What the agent is told before every conversation	`instructions`
Variables	`{{name}}` in the instructions, filled in when a run starts	`variables`
Capability	Something an agent may do, switched on in the Builder	`capabilities[]`
Skill	Know-how written once and read by any agent that needs it	`skill`
Context file	Standing knowledge an agent always has, such as a glossary or a policy	`context`
Knowledge base	Documents an agent searches and cites	`collection`
App	A page an agent publishes, such as a report or a small dashboard	`artifact`
Architect	The assistant in the corner of every page, which can build agents with you	`platform-assistant`

Where it runs¶

In the console	What it is	In the API and spec
Chat	The web conversation with an agent	`conversation`
Channel	Slack, Telegram or Mattermost, where people reach an agent	`channel_bot`
Routine	An agent running on its own, on a schedule or an event	`trigger`
Run	One time an agent answered, with its cost and steps	`agent_run`
Environment	A named stage, such as production, pointing at one version	`environment`
MCP server	An external tool server agents call, connected once	`mcp_connection`
Remote agent	Another vendor's agent, reached over A2A, that agents delegate to	`external_agent`
Agent2Agent (A2A)	The protocol other systems' agents use to send an agent tasks	`a2a_enabled`
Sandbox	An isolated computer where an agent writes files and runs code	`sandbox`
Documents	The files agents keep for a person, attached or written, in their sandboxes	`workspace`

Who decides¶

In the console	What it is	In the API and spec
Organization	The company or team everything belongs to	`organization`
Group	A department or team that agents, skills and knowledge are shared with	`group`
Lead	A group member who manages who else is in it	`is_lead`
Approval	A tool call waiting for a person to allow it	`approval`
Budget	A spending limit that stops runs when it is reached	`budget`
Guardrails	Checks that redact or block what flows through a run	`guardrails`
Vault	Where keys and tokens are stored, sealed and never shown again	`secret`

Capabilities¶

The name the Builder shows, and the id an agent's spec binds. Each is described in Capabilities.

In the console	In the spec
Apps	`artifacts`
Ask the user	`ask_user`
Web browser	`browser_use`
Chat channel lookup	`channel_tools`
Charts	`charts`
Date and time	`clock`
Calculations	`code_execution`
Coding agent	`coding_agent`
Long conversations	`compaction`
Context	`context`
Past conversations	`conversation_search`
Guardrails	`guardrails`
Image generation	`image_generation`
Knowledge search	`knowledge`
Media offload	`media`
Memory	`memory_files`
Memory (mem0)	`memory_mem0`
Planning	`planning`
Work while tools run	`background_tools`
Sandbox	`sandbox`
Skills	`skills`
Subagents	`subagents`
Instruction reminders	`system_reminders`
Thinking	`thinking`
Tool output limits	`tool_output_limits`
Tool search	`tool_search`
Visuals	`visuals`
Read web pages	`web_fetch`
Web search	`web_research`