Inert, not “OK”
A run that calls itself successful without having called a tool or summarised anything is counted as inert. Half the day’s runs inert: an alert. It is the failure nobody sees, because nothing falls over.
Your agent ran 340 times yesterday. It did nothing.
It “finished” every time. Without calling a tool, without producing anything. A cron job that answers OK and does nothing, only more expensive. Here each run is reported in one request, each agent is judged against itself — inertia, cost, failures — and sensitive actions wait for a human to say yes.
Cost, compared to the right day
A loop of model calls shows up on the bill before it shows anywhere else. Today’s cost is compared to the same weekdays: if it doubles, you know the next morning, not at the end of the month.
“I need to send this quote.”
The agent files the action, you get an email, you approve or refuse in one click, and the agent learns the answer. No reply within the deadline: refused. An agent never does by default what a human has not confirmed.
What it costs to set up
One request after each run, another before each sensitive action. No SDK, no framework imposed: any agent, any model.
What we never keep
What you send us: the agent’s name, the tools it called, the numbers, a summary. Not the conversations, not the documents. The summary is yours to fill with what can be shared.
What it does not do
No agent builder, no memory, no connectors — for now. Tock watches what your agents do and holds the door on sensitive actions. The rest will come when it works.