How Kyvvu compares
A label on a prompt doesn't stop a DELETE.
Guardrails check whether a prompt or a response is safe: injection attempts, toxic content, sensitive data in a message. Kyvvu decides whether an agent's next action may run, given everything it already did.
At a glance
Guardrails and Kyvvu, side by side.
Guardrails: A filter on the text going into or out of a model, which classifies a message or rewrites it.
| Guardrails | Kyvvu | Why it matters | |
|---|---|---|---|
| What it judges | One message: a prompt or a response | One action: a tool call, a write, a send, a delete | Agents do damage through actions, not only through words. |
| What it knows | The message in front of it | The whole task so far, in order | A harmless step after a sensitive one gets the verdict it deserves. |
| What it does | Classifies or filters the text | Returns allow, warn or block before the step executes | The harmful action never runs, whatever the text looked like. |
| Predictability | Often a model's judgement | Deterministic: the same path gets the same verdict | A decision can be replayed and explained to an auditor. |
| Where it runs | Often a service your text is sent to | Inside the agent's process: no proxy and no extra network hop | Prompts and data stay in your environment. |
Descriptions of other kinds of tools are general and simplified. Individual products differ, so check what yours does.
The difference
Where they differ.
- 01
Words or actions
A guardrail can tell you a message looks safe. It can't stop an agent from deleting records or sending data out. Kyvvu checks the step itself.
message/action
- 02
One message or the path
Reading a customer record is fine. So is sending an email. Doing both in one task is a leak. Kyvvu's rules can see that sequence.
path-aware
- 03
Rules people can read
Policies are YAML in your own git, reviewed like code, and every verdict names the rule that decided.
policies as code
Questions
Keep both?
01Do we still need guardrails?
Probably. They do a real job on the content of messages. Kyvvu covers what they can't see: the action and the path that led to it.
02Can Kyvvu replace a guardrail?
Not for content moderation. Kyvvu doesn't classify text; it decides whether a step may run.
See the difference
Watch the engine decide on your path.
In your browser
The playground
Build a task path and watch each step get a verdict.
For your team
Request a demo
A 30-minute demo with someone who builds Kyvvu, on the agents you run.