Most products bolt a chat box onto an existing dashboard and call the result an AI agent. Subscriber Bot took the opposite route, because the things an agent would touch here are recurring payments: a wrong move does not produce a bad paragraph, it cancels your insurance. So the first thing built was the record — what an agent is, who owns it, and precisely how much latitude it has — before any reasoning was wired to it.
Open AI Agents in the app and you get a roster. Each agent carries one of nine kinds, so its remit is legible at a glance: a Discovery agent hunts for recurring spend you have lost track of, a Renewal agent watches dates, a Cost agent watches money, a Negotiation agent is scoped to price conversations. A fleet-posture donut and live counters sit above the roster and are computed from your actual agents — how many are Disabled, how many Suggest-only, how many have run in the last seven days — never from placeholder numbers.
Posture is the important field. Every agent is created Suggest-only by default, and you can change any agent's posture inline from its card without leaving the list. Suggest-only means the agent may propose and never act; Disabled means it does nothing at all. The third posture, Autonomous, exists in the schema and is deliberately refused today — see "Autonomy that cannot be switched on by accident" for why that refusal is a feature and not a bug.
Open an agent and you see the policies that bound it. A policy is its own record with a condition and one of seven actions — Alert, Create task, Renew, Cancel, Pause, Upgrade, Downgrade — plus an enabled switch, a trigger count, and a last-triggered timestamp. You author policies from the agent's detail screen. A run-history timeline is derived from the agent's real last-run time and each policy's real last-triggered time, so an agent that has never run says exactly that rather than inventing activity.
What is honestly not here yet is the evaluator: the component that reads a policy's condition against your portfolio on a schedule and decides whether to fire. That engine is in progress, and until it lands, policies are authored, versioned, and visible — but nothing evaluates them on your behalf. The design order was intentional. The vocabulary of consent — who owns this agent, what may it do, under what rule, and what did it do last — is in place first, so that when the reasoning layer arrives it plugs into governance rather than the other way round.
Do it yourself
Stand up an agent and bound it — read the fleet posture, create a Suggest-only agent, change its posture from its card, add a policy on its detail page, and watch the fleet mix recompute.
Open AI Agents. The panel above the roster shows how your agents are allowed to act — the posture mix, the total, and how many ran in the last seven days.
You should see: You see, computed from your real agents, exactly how much latitude your fleet currently has.
Ready to make this your story?
