Mention tokens in agent instructions

Agent instructions can embed mention tokens — short references to org resources — that the platform resolves at run time. The syntax is:

/type[key]

The resolved resource is appended to the agent’s system prompt as a “Referenced Resources” block that tells the agent exactly what the resource is and which tool to use to act on it. The agent sees the friendly description; you just embed the token.

Supported types

Token	key value	What the agent is told to do
/prompt[slug]	Prompt template slug	Call run_prompt with prompt="<slug>" and the required variables (or material input, if the prompt has an extraction schema)
/agent[slug]	Agent slug	Start a work chat via create_chat with agent_slug="<slug>" (fire-and-forget)
/skill[slug]	Skill slug	Call activate_skill with slug="<slug>" to load its instructions
/tool[name]	Custom tool name	Call the tool directly by name
/connected-tool[slug]	Integration slug (or name)	Use the integration’s tools (available prefixed with <slug>_)

A prompt with an extraction schema (fields, see prompt-fields) is still referenced with /prompt[slug] — there is no separate extractor token.

Example

An instruction that classifies a ticket, drafts a reply, and optionally escalates:

When a new support ticket arrives:
1. Use /prompt[ticket-classifier] to extract the category and priority.
2. If priority is "urgent", hand it off to /agent[escalation-bot].
3. Otherwise, use /prompt[draft-reply] with variables {category} to draft a reply.

Notes

• Keys are slugs (not display names) for prompts, skills, and agents. Use list_prompts, list_skills, or list_agents to find slugs.
• Custom tool keys are the tool name (not slug), matching the name it was created with.
• Unresolved mentions (typos, deleted resources) are handled gracefully — the agent is warned in-context that the resource was not found.
• Duplicate tokens are deduplicated; order of first appearance is preserved.