For risk and internal audit
Test the control on every run, not a sample.
Agent risk doesn't fit frameworks built for static models and for people. Kyvvu gives you controls that operate on each step an agent takes, and a record that shows they did.
The problem
Agents don't fit the control framework.
- 01
Frameworks made for models
Model risk management validates a model before it goes live. An agent's behaviour depends on its prompt, its tools and what it reads at runtime.
model risk management
- 02
Sampling can't keep up
An agent can take thousands of decisions a day. Testing a sample tells you little about the run that goes wrong.
control testing
- 03
Design isn't operation
A documented control shows design effectiveness. You need evidence that it operated, on each run, for the whole period.
operating effectiveness
In practice
Controls that leave evidence every time they act.
A payment needs a second pair of eyes
A finance agent prepares a payment. The control: a person approves before any financial write.
Illustrative scenarios. Each one is decided by a single policy from the open policy library, as the engine evaluates it. References to laws and regulations are our reading of them, to show how a policy maps to an obligation. Not legal advice.
How Kyvvu helps
Evidence of operation, produced by the control itself.
- 01
A record of every step
Each task is recorded as an ordered trace of its steps with the verdict on each, hash-chained when it reaches the platform.
logs/traces
- 02
The same answer every time
Verdicts are deterministic: the same path gets the same verdict, so a decision can be replayed and explained.
deterministic verdicts
- 03
Every change on record
Policies live in git, so you can see which rule applied when, who changed it and who approved the change.
versioned policies
- 04
Incidents your team can follow up
Every warn or block opens an incident with the policy, its severity, the step and the agent's risk level.
incidents
Introduce a control safely
Observe, then flag, then enforce.
Measure a control on real traffic before you rely on it, with the same policy file all the way through.
Shadow
Measure
Every verdict is recorded and nothing is blocked. You see how often the control would fire.
step.resource POST payments/transfers, without approval
ALLOWruns · logged as a would-be block
Suggest
Flag
Risky steps still run but get a WARN, so owners can tune the rule without stopping work.
step.resource POST payments/transfers, without approval
WARNruns · flagged to the owner
Enforce
Prevent
The step is blocked before it runs, and every verdict is on record.
step.resource POST payments/transfers, without approval
BLOCKstopped before it ran
Questions
Common questions.
01Can we replay a decision?
Yes. Verdicts are deterministic: the same path under the same policy version gets the same verdict, so you can re-run a task's path and check the outcome.
02Is the record tamper-evident?
Traces are hash-chained when they reach the platform, so a removed or changed step breaks the chain.
03Can we see which agents carry the most risk?
Each agent registers with a risk classification, and every incident carries the agent's risk level, so you can follow up on the high-risk ones first.
Get started
Every agent has to ASK before it acts.
Start with one agent
See it on your own agents
A 30-minute demo with someone who builds Kyvvu: your agent, your rules, and the policy that decides each step.
Rehearse it
An agent incident tabletop
Test your agent incident process in 60 minutes and get a scored after-action report.
Want to see other teams and sectors? All use cases.