AI agent guardrails
Most guardrails work on text. The guardrails that matter for a refund, a discount or a deletion work on actions.
What are AI agent guardrails?
AI agent guardrails are the limits placed on what an AI agent can do. Prompt level guardrails shape the model output. Action level guardrails run outside the model and decide whether a specific business action may execute.
Three layers of guardrails
- Model level: system prompts, instructions and output filters
- Access level: which tools and scopes the agent holds
- Action level: whether this specific action, right now, is within policy
Why the first two layers are not enough
An instruction is a suggestion to a probabilistic system. It can be followed 99 times and missed once. When the action is a payment, once is enough to matter.
Tool access is binary and static. An agent that can call the refund endpoint can call it for 20 EUR or 20,000 EUR.
Action level guardrails close that gap. They read the actual system state and apply the rule that the business already follows for humans.
What good action level guardrails look like
- Deterministic: the same input always produces the same decision
- Evidence based: decisions cite the records they used
- Explainable: every decision comes with a reason
- Fail safe: missing evidence returns INSUFFICIENT EVIDENCE, not a guess
- Reviewable: policies are written in business language, owned by the business
Questions
Frequently asked questions
Give your agents freedom without giving them a blank cheque.
Tell us what your agents are allowed to do today. We will review whether Gatewise fits your agent workflow.
Works with your agents and existing business tools