Skip to main content
An agent persona is a catalog resource that defines a reusable template for agents. When you spawn an agent with --agent=<name>, it inherits the persona’s system prompt, model, tool configuration, and default tasks. Agent personas are tenant-scoped. Platform-managed agent personas exist in every tenant. Custom agent personas are authored via murmur set agent-persona. An agent persona accepts input in two formats: a Markdown file with YAML frontmatter (the content field), or individual structured fields (description, prompt, model, etc.). When structured fields are present, they are the source of truth and any content in the request is ignored — the server re-renders content from the structured fields on every write. Provide content alone to author the persona as Markdown.

Fields

AgentTask fields

Each entry in the tasks list has:

Follow-up overrides

When an event wakes an agent — a PR comment, a PR review, a CI result — Murmur delivers two things: the content (the platform-formatted facts: who, where, the comment body, the exact gh commands) and the instructions (the prose telling the agent what to do about them). A persona can override the instructions and the checklist tasks for a given event class. The content is always platform-formatted, so you never re-derive event formatting. Author overrides as the followups: frontmatter map, keyed by event class. At most one override per class. Each override has: Declaring an override for a class replaces that class’s default tasks entirely. Classes with no override keep their default instructions and tasks.

Input formats

Markdown with frontmatter

Pipe a Markdown file where the YAML frontmatter contains the structured fields and the body is the system prompt.
Frontmatter keys use camelCase: disallowedTools, maxTurns. This matches the format that the server renders on read.

Structured fields

Post individual fields as YAML. The prompt field carries the system prompt directly.
Do not combine content with structured fields (description, prompt, model, etc.) in the same request. The server rejects requests that provide both.

Examples

Creating an agent persona

Listing agent personas

Platform-managed agent personas are marked [platform].

Reading a single agent persona

The response is the server-rendered Markdown — YAML frontmatter followed by the system prompt body.

Spawning an agent with a persona

The agent inherits the agent persona’s prompt, model, tools, and default tasks.

Tasks with activation conditions

Plain strings become tasks that are always active. Structured entries with activate_when activate only when the agent pushes changes matching the glob pattern.

Overriding follow-up instructions

A pr_comment follow-up now delivers the platform-formatted comment, then these instructions, and seeds the one listed task instead of the default. pr_review has no override, so it keeps its default instructions and tasks.

Deleting an agent persona

Platform-managed agent personas cannot be deleted.

Errors