The agent can touch real campaigns, so it follows conservative rules by default.
The safety defaults
New campaigns launch paused. When the agent creates campaigns or ads, they are created in a paused state unless you explicitly say otherwise. Nothing spends because of a chat message you did not mean that way.
Evidence only. The agent reports numbers it actually pulled from your accounts. If it could not get the data, it says so instead of inventing a plausible answer.
Reversible first. When it recommends changes, it leads with reversible, low-dependency moves before anything drastic.
Channel specialists. Questions about Meta, Google or LinkedIn are handled by a dedicated sub-agent for that platform, which only appears once you have connected the channel.
Models
The model picker in the input bar lets you choose the AI model behind the agent. The default is a strong general model; switch if your team prefers a specific one. All models follow the same rules and use the same tools.
What it remembers
The agent keeps working memory inside each thread, and writes durable knowledge to the client's Brain, which you can read and edit at Library > Brain. That is what "you never have to explain it twice" means in practice: corrections you make become part of how it works with that client. See The Brain.