Sample recordA day of agent traffic — none of it real, all of it verifiable.Make it real — connect your agent →

Rules

What is armed right now, and what each rule actually stops.

Sample data⏻ halt
Add a rule

Active rules

6 of 6 enforcing
When the kill switch is on, no agent may act at all.
kill-switch1
EnforcingBuilt in

Tenant-level halt. Forbid-wins over everything.

Moving more than the threshold requires a named human to approve first.
human-signoff-above-threshold1
EnforcingBuilt in

EU AI Act Art 14 human oversight; financial supervision.

A decision against a person must record why, in structured form.
adverse-decision-requires-reason1
EnforcingBuilt in

Colorado SB 26-189 — adverse decisions carry a reason and a reviewer.

An agent may only use the tools it was explicitly granted.
tool-allowlist1
EnforcingBuilt in

Deny-by-default tool scope, bounded by the invoking user's own rights.

An agent spending fast enough to blow its daily ceiling inside the hour is stopped now, not at the ceiling.
burn-rate1
EnforcingBuilt in

The runaway-loop guard. A ceiling reports the damage; burn rate stops it in the first minute.

An agent that exceeds its daily ceiling or loops too long is stopped.
cost-velocity1
EnforcingBuilt in

Runaway-loop and spend guard. The FinOps rule that is also a compliance control.

On your own log you write rules here in plain English backed by a condition tree — and every version is kept, because a rule that governed a decision is part of that decision's evidence.

Every rule can run in monitor mode, which records what it would have done without acting. The arming sequence is record → warn → enforce: a rule that starts blocking production traffic the day it ships does not survive its first false positive.