Keeping AI agents in check: an overview

Always available

• Guardrails (Project settings → Guardrails): approved claims, banned phrases, required disclosures and brand rules, and whether a broken rule warns or blocks agent publishing. A project with no rules is never flagged. [Read more](/dashboard/help?article=agent-guardrails)
• Agent activity (Project settings → Agent activity): every change an agent made, and every publish your rules stopped. Agencies can show a simple version to clients in the portal. [Read more](/dashboard/help?article=agent-activity-log)
• Brand rules (inside Guardrails): plain-language guidance every agent follows when it writes. [Read more](/dashboard/help?article=brand-rules)
• The knowledge base keeps up: approved claims, brand rules, everything that passes the approval chain and every fact change are written to the project’s knowledge base automatically, so agents, knowledge search and the claim check all use them. [Read more](/dashboard/help?article=knowledge-maintained-documents)

Off until you turn them on

• Learn from feedback: when someone rewrites or deletes an agent’s draft, you say why once and agents follow it from then on. [Read more](/dashboard/help?article=learn-from-feedback)
• Agent permission roles: let each kind of agent do everything, only prepare drafts, or only read. [Read more](/dashboard/help?article=agent-permission-roles)
• Approval chain: Brief → Copy → Creative → Legal → Launch, with owners for each step; nothing an agent made goes live until every step is signed off. [Read more](/dashboard/help?article=approval-chain)
• Learn on-brand from past work: agents study your approved, team-rewritten and published work before writing, and reuse your brand images. [Read more](/dashboard/help?article=brand-learning)
• Change a fact everywhere: change a price or legal line once; drafts, posts, ads, the knowledge base and guardrails follow. [Read more](/dashboard/help?article=fact-changes)
• Creative performance breakdown: which hook, offer, call to action and image drove each ad’s results. [Read more](/dashboard/help?article=creative-breakdown)
• Coming next: project tools (Asana, Linear, Notion), also off until you turn it on.

What happens when an agent publishes

1. 1Permission roles (if on): is this agent allowed to publish on this project at all?
2. 2Approval chain (if on): is the item approved, unchanged since it was approved?
3. 3Guardrails: banned phrases and required disclosures are checked; with “Block”, a broken rule stops the publish.
4. 4The result, published or stopped, is written to Agent activity.