> ## Documentation Index > Fetch the complete documentation index at: https://docs.clarion.cantina.xyz/llms.txt > Use this file to discover all available pages before exploring further. # Agents > The read-only system prompt every triage starts from, and the settings you layer on top of it: agent behavior, escalation and notifications, and skills. **Settings → Agents** shapes how every agent in the workspace triages. It opens on the system prompt, so you can see what agents are told before you change anything. The page has four sections, each linked from the settings sidebar under **Agents**: | Section | What it holds | Editable | | - | - | - | | [System prompt](#system-prompt) | The default prompt every triage run starts from | No | | [Agent behavior](#agent-behavior) | Standing instructions every agent follows | Yes | | [Escalation & notifications](#escalation-and-notifications) | Severities that escalate, and where agents send escalations, approvals and questions | Yes | | [Skills](#skills) | Playbooks agents open on demand | Yes | Changes take effect on the next agent run. Drafts are kept when you switch between sections. ## System prompt The default prompt Clarion gives every triage agent in the workspace. Clarion maintains it and you can't edit it; the version shown next to it changes whenever Clarion updates the prompt. What you *can* change is added on top of it: **Agent behavior** and **Escalation & notifications**. The escalation bar is written into the prompt itself, so the text you see reflects your current setting. Skills are not part of the prompt. Each agent sees the list of skills assigned to it and opens one only when an issue calls for it, so skills shape how agents work without making every triage longer. ## Agent behavior Standing instructions every agent run inherits, on top of whatever its skills say. These were previously called *workspace instructions* and lived under **Settings → Workspace**. Over [MCP](/learn/settings/mcp-access), read and replace them with `get_agent_behavior` and `set_agent_behavior`. Use them for context that doesn't belong to any single skill: * Your business context, and who owns which system. * Organization-wide terminology, escalation hierarchy, and on-call conventions. * Environment-specific quirks ("staging fires this alert nightly; it is expected"). * What agents must never do. Notification channels do not belong here. Configure them under [Escalation & notifications](#escalation-and-notifications); agents read that set directly and pick a destination per run. The editor supports the same `@`-references as the skill editor, so instructions can point at a specific skill by name. Keep instructions short and high-signal. They are added to every run and compete with skill content for the agent's attention. If guidance only applies to one class of issue, it belongs in a skill, not here. ## Escalation and notifications Decides when an agent pages your team, and where it sends escalations, approval requests and questions. Destinations, routing and the on-call rotation are covered in [Notifications](/learn/notifications/introduction). ### Severities that escalate When an agent finishes triaging an issue it considers a real problem, it **escalates** it: the issue moves to `Escalated` and the team is paged. This setting is the bar for that page. The agent escalates only when the issue's severity is one you selected. Anything below the bar moves to `Open` instead, so it still waits for a human, just without a notification. * **Default:** `High` and `Critical`. * **Choices:** any combination of `Low`, `Medium`, `High`, and `Critical`. * **Minimum:** at least one severity. The last one can't be unchecked. A few things are worth knowing: * **The bar applies to agents only.** A member who clicks **Escalate** on an issue always escalates it, whatever its severity. * **Parking an issue in `Open` doesn't change the verdict.** Its disposition and severity stay as triage set them; only the status differs. * **Some issue kinds are stricter.** Issues of kind **Risk** or **Vulnerability** stay `Open` on an agent's decision regardless of this setting. The one exception is a `Critical` finding the agent justifies as imminent danger. See [Issue kinds](/learn/issues/introduction#issue-kinds). * **Agents are told the bar.** Your selection is written into the [system prompt](#system-prompt), so the agent works from your policy rather than the default. Don't restate or override the escalation bar in skills or agent behavior, for example "escalate every Medium finding". An instruction that conflicts with the system prompt confuses the agent, and the setting still decides the final status. To change which severities escalate, change this setting. Adding `Medium` pages more often and catches more. Narrowing to `Critical` alone keeps pages rare but leaves more for your team to sweep in the `Open` queue. Start from the default and adjust once you see your volume. ## Skills Step-by-step playbooks agents open on demand for specific kinds of alerts, such as a phishing report or a suspicious sign-in. Each row shows the skill's description, which is the only text an agent uses to decide whether to open it; missing descriptions and broken references are flagged. See [Skills](/learn/skills/introduction) and [Authoring best practices](/learn/skills/authoring-best-practices). Domains, human-input and Teams scoping toggles, and BYOK. Destinations, routing, and the on-call rotation.