Personality & Guardrails

Personality and guardrails are the foundation of every Rival Agent. Personality (sometimes called persona) defines how an agent communicates; guardrails define what it’s allowed to do. Together they’re the difference between an agent that’s helpful and on-brand and one that’s a liability.

Personality

An agent’s personality is built from three optional elements:

• Tone — its communication style (e.g., “direct and focused,” or “warm and conversational”).
• Working style — how it approaches tasks (e.g., “proactive — proposes a plan before acting”).
• Values — behaviors it should consistently follow or avoid (e.g., “always cites sources for factual claims”).

When an agent is published, these details appear on its listing so potential adopters understand how it behaves.

Guardrails

Guardrails assign a permission level to specific actions:

• Allowed — the agent performs the action freely.
• Needs approval — the agent asks for human confirmation before proceeding.
• Blocked — the agent refuses the action entirely.

Common examples: allow reading from a data source, but block sharing customer PII or modifying a production database.

The enterprise layer: org-wide guardrails

Here’s what self-serve doesn’t have. In Rival Enterprise, an admin can set organization-wide guardrails that apply on top of every agent’s own — a floor no individual agent or user can drop below. Even if someone builds an agent and forgets a limit, the org-level guardrails still hold. See Guardrails & allowed models.

Guardrails during rituals

When an agent runs on a schedule (a ritual), there’s no human in the chat to approve anything. Actions marked Needs approval are skipped during autonomous runs, with a note added to the log. Anything you want a ritual to actually do should be marked Allowed.

A quick example

An analyst builds an agent to answer questions about customer accounts. Its tone is factual and concise; its guardrails block exposing full account numbers. On top of that, the organization’s guardrails forbid any agent from sending data to an unapproved connector — so the analyst’s agent is safe by default, without the analyst having to anticipate everything.