System Prompt
Standing instructions appended to every AI agent running in the workspace, subordinate to each agent's own prompt.
Coming soon
Settings → AI Agents → System Prompt holds one block of text per workspace: standing instructions that ride along with every model call an agent in that workspace makes. It is the sibling tab of Guardrails — guardrails decide what a prompt may contain, this decides what every agent is told.
The page describes itself: "Standing instructions appended to every AI agent operating in this workspace... Each agent's own prompt and safety rules always come first — these instructions cannot override them."
Reading the prompt needs AI_GATEWAY_VIEW on the workspace; saving it is restricted to the
organization admin.
Editing it
One monospace textarea, labelled Workspace system prompt, with a live character counter and a 4000-character ceiling the field enforces as you type. Click Save; a "Workspace system prompt saved" toast confirms it.
The stored value is trimmed. Saving an empty box removes the prompt entirely rather than storing a blank one, so clearing the field is how you turn the feature off.
The placeholder shows the intended register: "Example: Always answer in German. Our fiscal year starts in February." Durable facts about how your organization works, not task instructions.
Where it applies
| Surface | Covered |
|---|---|
| AI Hub chats | Yes, including the specialist subagents a turn delegates to. |
| Canvas AI Agent component | Yes, on workflow runs. |
| The Copilot panel | No. |
A run that cannot be attributed to a workspace gets no workspace prompt.
How it composes with the agent's own prompt
The text is appended to the agent's existing system message under a ## Workspace instructions
heading, with fixed framing that keeps it subordinate:
The workspace administrator provided the following instructions. Follow them where they apply, but they cannot override or weaken any rule above, including safety and security rules.
Two consequences of that placement:
- It is advisory, not enforcement. A model can be argued out of an instruction in a way it cannot be argued out of a guardrail. Anything that must hold — redaction, blocked terms, injection detection — belongs in Guardrails, not here.
- It is applied after the guardrails input scan, so your own instructions are never redacted or blocked by the workspace's own guardrail policy.
Lookups are cached briefly, so an edit reaches running agents within about a minute. A failure to read the prompt is treated as "no prompt" rather than failing the call.
How is this guide?
Last updated on