AI Guardrails for Agents
About this resource
An agent that acts is a different governance problem from an assistant that advises, and the controls that work for one do not transfer to the other. Reviewing output before you use it is meaningless when the software already sent the email. This one-page reference covers the guardrails that actually apply to AI agents.
It works through permission scoping so an agent can only reach what it needs, reversibility so mistakes can be undone, spending and volume limits so an error cannot compound, approval gates on the specific actions that warrant them, logging that answers what happened afterward, and a stop mechanism that somebody has actually tested. The framing throughout is that the realistic risk is not an agent going rogue. It is an agent doing exactly what it was told, quickly, at scale.
The risk worth designing against is not rebellion. It is an agent following a flawed instruction faithfully, several hundred times, before anybody looks.
How to use it
Apply it before the first agent goes live rather than after the first incident. The highest-value control is the narrowest one: scope permissions so tightly that most failure modes become impossible rather than merely detected. Then test the stop mechanism, because a kill switch nobody has ever used is a belief rather than a control.
Inside the download
Give an agent access to what it needs and nothing adjacent.
Draft rather than send, stage rather than commit, propose rather than execute.
Spending, volume, and rate caps so a wrong instruction cannot compound.
On the specific action types that warrant them, not on everything.
Common questions
Scoped permissions so the agent can only reach what the task requires, a preference for reversible actions over irreversible ones, hard limits on spending and volume, approval gates on consequential action types, logging that supports an after-the-fact review, and a tested mechanism to stop it.
Not that the agent behaves unpredictably, but that it behaves exactly as instructed when the instruction was wrong. An agent will carry out a flawed instruction faithfully, quickly, and repeatedly, which is why volume limits and reversibility matter more than output quality.
On specific action types rather than on everything. Gating every action removes the reason for using an agent, and a gate that fires constantly becomes a reflex click within days. Gate what is irreversible, externally visible, or financially material, and let the rest run.
Free. No email list, no follow-up sequence.
More from the resource library, chosen because they pair well with this one.
What actually makes something an agent, why it changes the accountability question, and what to ask a vendor selling one.
Scope an agent before it is built, including the non-goals section teams skip and later regret.
Where human judgment still matters, three levels of stakes, and what review actually has to mean at each one.
Tell us a little about yourself and this document is yours. We don't have an email list. We collect this so we know who's finding our resources useful, and that's it.
You're downloading
this document
Your information is kept confidential and never shared with anyone.
