Kyvvu

Runtime policy enforcement for AI agents

Your organization has rules about customer privacy, data handling, decision transparency, and appropriate escalation. But when you deploy AI agents, how do you ensure they actually follow these rules at runtime?

The problem: AI agents make decisions autonomously. You can prompt them, but you can't guarantee they'll always do the right thing. One bad decision—exposing PII, approving the wrong transaction, bypassing required approvals—can cost you. And when something goes wrong, can you prove what happened?

What Kyvvu does: We enforce your organization's policies at runtime, across every agent, regardless of framework. Define a rule once — in plain language, mapped to your EU AI Act obligations — and Kyvvu enforces it automatically at the code level, between every agent step. Not after the fact. Not optionally. At the moment the agent acts.

How it works

Kyvvu captures every step your agents take and evaluates them against policies you define:

from kyvvu import Kyvvu

kv = Kyvvu(api_key="your-key")

# Register your agent once
kv.register_agent(
    agent_key="customer-support-agent",
    name="Customer Support Agent",
    purpose="Handles customer inquiries",
    risk_classification="HIGH"
)

# Decorate functions to log automatically
@kv.log_step("LLM_CALL")
def generate_response(query):
    return llm.invoke(query)

@kv.log_step("TOOL_CALL", has_write_permission=True)
def update_customer_record(data):
    return database.update(data)

Three core capabilities:

• Policy enforcement — Rules defined once, evaluated at runtime across all agents, at registration and between every step — regardless of framework
• Immutable audit trails — Hash-chained logs at agent, task, and step level — tamper-evident and audit-ready
• Incident management — Automatic incident generation when policies are violated, with full context for investigation

We integrate seamlessly with LangChain (via callback handlers), work with any Python framework, and deploy on your infrastructure. Framework-agnostic REST API for non-Python agents.

Where we are now

What exists today:

• Production-ready REST API and Python SDK with decorator-based integration
• LangChain callback handler for automatic monitoring
• 23 policy templates aligned with EU AI Act requirements, mapped to specific articles
• Dashboard showing agent activity, policy violations, and incidents
• Hash-chained immutable audit logs for compliance evidence
• Structured incident reporting with full context

What we're building:

• Large-scale testing (currently running internal evaluations)
• Integration with existing incident response workflows
• LangSmith/Azure/AWS log ingestion for platform-native monitoring

Want early access? Email maurits@kyvvu.com.

About Kyvvu

Kyvvu is founded by Maurits Kaptein. Kyvvu is an early-stage company building runtime compliance infrastructure for AI agents. We're backed by seed funding and working with a small number of pilot partners in regulated industries across the Netherlands.

Learn more about us or see open positions.