Articles · Money & Agents
Your own agent, at three in the morning
Identity stops intruders. It does not stop a properly identified agent that starts behaving badly. Behaviour Watch reads actions against the mandate as they happen — and can stop an agent the moment it steps outside.
By Lucian · Lux Automations
The problem
Identity says an agent may enter. It says nothing about what that agent does once it is in — and the dangerous agent is usually your own. A properly identified agent can misread its instructions, fall into a loop, or be steered by poisoned input in the material it reads, all while presenting a perfectly valid identity at every door.
Agents also work unsocial hours. They act at 3 a.m., while the people responsible for them sleep. A dashboard alert that a human reads at 9 a.m. is not a control on an agent that acted six hours earlier — it is a report on damage already done. Watching behaviour only in hindsight is not watching it at all.
The technology
Behaviour Watch reads an agent’s actions against its mandate continuously, as they happen — not in a review afterwards. The mandate is the owner’s own statement of what this agent may ask, send, spend and set in motion, so “behaving badly” has a concrete meaning: an action that falls outside those lines.
The moment an action falls outside the mandate, the crossing bounces and the agent’s identity is suspended, pending a human. Where the checks disagree, the guard stops and calls a person in — it never quietly settles things for itself, and no agent decides a dispute it is part of. Alongside the automatic watch sits a kill switch the owner holds: pull it at any time, for any agent, and that agent stops instantly while the others keep working. Every step leaves a signed receipt, so the watch itself can be audited.
What it changes
What changes is the hour at which problems are met. Instead of discovering at breakfast what an agent did overnight, the owner is called in at the moment an action steps outside the mandate — with the agent already stopped, and a receipt showing what it attempted.
That makes it safer to let agents do real work at all. An owner who knows a watcher will catch the first step outside the lines, and that they hold a switch of their own, does not have to choose between locking agents away from useful work and lying awake wondering what they are doing.
The watch is deliberately not a judge. It does not decide that an agent meant well or meant harm, and it does not hand out second chances on its own authority. It compares an action with a mandate its owner wrote, stops the action that falls outside, and puts the question in front of a person, with the receipts attached. That modesty is the design: software is good at noticing the first step outside a line, and people remain the ones who decide what the step means.
See it
Behaviour Watch + Kill Switch watches the agent you already trust at the door, for the whole time it is inside. Behaviour Watch + Kill Switch — the product page · All the house articles · Lux Automations — ailux.ca