The problem this solves
Autonomy and spend control pull against each other, and most systems resolve the tension badly in one of two ways. Either every action needs a click, which buries the owner in confirmations until approving becomes reflexive and meaningless. Or the system runs free within some monthly cap, and the owner discovers at review time which expensive missions ran, after the money is spent and the choices are unexaminable.
The practical failure is location. Approval requests traditionally live in a console, and the person who must approve is in an airport, a meeting, or a car for most of the working day. So requests queue against a laptop that will not open for hours, urgent work stalls behind an absent click, and eventually someone widens the autonomy just to stop the stalling - which reintroduces the original problem at a higher price.
A workable gate has three properties. It triggers on projected cost rather than on action type, so routine work flows and only expensive work asks. It reaches you on whatever surface you are actually on, phrased so the decision takes ten seconds: what it costs, what you get. And its default is refusal - silence spends nothing - so an unread message can never become an authorized expense.
How the mission runs
- You set one number: the threshold. Any mission whose projected credit cost crosses your threshold must ask before running; anything under it proceeds normally. One number defines the boundary between routine autonomy and checked spend, and you can move it anytime as trust in the system grows or a tight month demands a lower bar.
- The estimate is computed before work begins. When a mission is composed - by you in chat, or by a schedule firing - the agent projects its credit cost against the threshold first. Crossing missions pause at zero spend. The gate sits before execution, so there is no partial work to unwind when the answer is no.
- The ask arrives where you started the mission. The request goes to the surface the mission came from - the Slack thread, the Telegram chat, or your email via the Email Agent - phrased for a fast decision: what the mission does, the credit estimate, what you get back. One message, one decision, no console required.
- Your yes releases it; silence never does. An explicit approval in the thread starts the run, and the reply confirms it is underway. No response leaves the mission paused indefinitely at zero cost. Expiry after your configured window closes the request cleanly, so a message you never saw cannot ripen into spend.
- Every gate decision is on the record. Approvals, declines, and expiries are logged with the estimate shown and the thread where the decision happened, and actual cost meters against the same workspace balance as everything else. At review time you can see each gated mission, who released it, and what it truly cost beside its projection.
The prompt
This is the exact objective the agent receives. Swap the obvious placeholders for your own domain, segment or channel and run it as-is from the console, Slack, or the API.
What comes back
A spend gate that travels with you: expensive missions pause and ask on the surface where you started them - Slack, Telegram, or email - with a credit estimate and the expected payoff in one message. Approved work runs immediately and confirms; ignored requests spend nothing, forever. Underneath, a clean log of every ask, every decision, and projected versus actual cost.
Make it yours
- Set a second, higher threshold that additionally notifies a teammate, so routine overages clear on your yes while unusual ones get two sets of eyes.
- Lower the threshold to zero for one experimental week so every mission asks, as a fast way to learn the true cost shape of your workload before setting the real bar.
- Route approvals by mission type: research overages ask in Telegram where you answer fast, while anything touching an external system asks in the team's Slack channel.
Frequently asked questions
What happens if I simply never answer?
Nothing runs and nothing is spent. Holding is the default, and it costs zero: the mission stays paused with its estimate attached until you approve, decline, or your configured expiry closes it. There is no timeout that converts silence into consent anywhere in the gate.
How accurate are the credit estimates?
Estimates are projections from the mission's composed steps and typical costs for each, stated as estimates rather than quotes. The log records actual beside projected for every gated mission, so calibration is visible over time, and consistent divergence is a signal to tighten how missions get scoped.
Does approving in Telegram behave differently from approving in Slack?
No. The surfaces are doors into the same workspace: one meter, one approval record, one mission state. An approval from any connected surface releases the same run, and the decision is logged identically. The gate follows the mission rather than the app you happened to be in.