What is a guardrail?
A hard limit an AI agent cannot talk itself past, enforced by the system that runs the action.
Guardrail — A guardrail is a fixed rule an AI agent cannot override — a daily send cap, a minimum gap between touches, a business-hours window, a confidence floor. Guardrails bound what an agent in auto mode can do without asking.
Why it matters
An agent in auto mode is going to take actions nobody reviews before they happen. A guardrail is what keeps those actions inside a range a person already agreed to. A daily send cap stops a runaway loop from emailing a thousand contacts overnight. A minimum gap between touches stops a prospect from getting three messages in one hour. A confidence floor routes the agent's unclear cases to a person instead of letting it guess.
The guardrail has to be enforced by the code that executes the action, not by the instructions given to the model. A prompt that says "never send more than fifty emails a day" is a request; a model under enough pressure from its own reasoning, or fed a malformed instruction, can ignore it. A cap enforced by the sending service itself cannot be talked past, because the fiftieth email simply does not go out. Guardrails are usually set once by an admin and apply to every run of an agent, regardless of what that specific run decides is a good idea.
How a guardrail is applied
- 1
A limit is defined
A daily cap, a time window, a minimum gap or a confidence floor, set by a person.
- 2
It is enforced in the executing system
The sending service, scheduler or database checks it, not the model.
- 3
The agent proposes an action
It decides what to do based on its own reasoning over the record.
- 4
The system checks the guardrail before running it
An action that would break the limit is blocked or deferred.
- 5
A blocked action is logged
So a person can see how often the guardrail is actually doing work.
The mistake to watch for
Questions
- How is a guardrail different from a policy?
- A policy can be a broader statement of intent that still needs a human to apply judgment. A guardrail is narrow and mechanical: a specific limit checked by code before an action runs, with no judgment involved.
- Can a guardrail be changed by the agent itself?
- No. A guardrail that the agent it constrains can edit is not a guardrail. Changing one should require an admin action that leaves its own audit trail.
- What happens when an agent hits a guardrail?
- The blocked action is either dropped, deferred to the next allowed window, or routed to a person for review, depending on how the guardrail is configured.