Skip to main content
An agent persona is a reusable template for how an agent should think and what tools it should use. It’s the difference between spawning “a Claude” and spawning “a code reviewer” or “an architect drafting a proposal” — same model underneath, different instructions, tool budgets, and behavior. Personas live in the catalog as the agent-persona kind. They are tenant-scoped, version-controlled, and authored in Markdown with YAML frontmatter or as structured YAML.

​
What a persona contains

The prompt is the heart of the persona. Write it as you would a senior engineer’s briefing: principles, conventions, what to avoid, what to prefer.

​
Follow-up instructions

When an event wakes an agent — a PR comment, a PR review, a CI result — the follow-up has two parts: the platform-formatted content (the facts of the event) and the instructions (what to do about them). A persona can tailor the instructions and the checklist tasks per event class through its followups map, so a payments reviewer can react to PR comments differently from a generalist while still receiving the same platform-formatted event. The content is always platform-formatted — personas shape only the response, never the event formatting. See the agent-persona reference for the format.

​
Built-in personas

Murmur ships platform-managed personas that exist in every tenant: Use them via the --agent flag on murmur spawn, or select them in the dashboard’s spawn dialog.

​
Authoring your own

Create personas with murmur set agent-persona — either piping a Markdown file with YAML frontmatter, or structured YAML. Both produce identical catalog records. See the agent-persona reference for all fields and formats.

​
Lifecycle

The agent inherits the persona’s tools, model, and system prompt at spawn time. Edits to personas affect future spawns only — running agents continue with the persona they were spawned with.