VaikoraVaikora

Vaikora › Blog › Product Updates

Deterministic policy enforcement for AI agents: what it means and how Vaikora implements it

Product Updates · Nov 17, 2026 · 4 min read

Probabilistic guardrails ask "how likely is this action to be harmful?" Deterministic enforcement asks "does this action match a policy rule?" and produces a binary, repeatable answer. The difference matters when an AI agent is about to write to a production database or call an external API with customer data.

Why probabilistic checks are not enough

Most AI safety layers rely on a secondary model to judge whether a primary model's output looks acceptable. The problem: the same input can produce different judgments on different runs. That non-determinism is fine for content tone. It is not acceptable for a policy that says "never send PII to an external endpoint" or "require human approval before deleting records."

A compliance auditor asking why a specific action was allowed or blocked needs a definitive answer. "The safety model thought it was probably fine" does not satisfy a SOC 2 audit or a GDPR investigation.

What deterministic enforcement actually means

Vaikora sits inline between an AI agent and every LLM call or tool execution. Before any action runs, Vaikora evaluates it against a policy file. The evaluation always produces one of four decisions: allow, block, modify, or require-approval.

Identical input always produces an identical decision. No sampling, no temperature, no model variance. The policy is code, and code is deterministic.

Every decision is written immediately into a SHA-256 hash-chained audit log. Each entry is signed against the previous entry, so the log is tamper-evident. If someone asks later why a specific tool call was blocked at 14:32:07 on a given date, the answer is in the log and the log cannot have been silently altered.

A concrete scenario

An AI agent is processing a support ticket queue. It has access to a CRM tool, an email-send tool, and a knowledge base search tool. A ticket arrives that contains a customer's payment card number in the body text.

Without a runtime gate, the agent might summarize the ticket, include the card number in its reasoning context, and pass that context to the email-send tool as part of a reply draft.

With Vaikora inline, the content inspection module catches the PCI-DSS pattern in the outbound payload before the email-send tool executes. The policy rule set to "block" fires. The agent receives a blocked response. The card number never leaves the system. The audit log records the decision, the matched rule, the timestamp, and the hash of the payload (not the payload itself).

The agent did not need to be retrained. The CRM tool did not need to change. The policy enforced the boundary at the moment of execution.

How the policy layer is configured

Vaikora's policy file is YAML. Rules reference content inspection modules, tool names, data classification labels, or combinations of all three. The six content inspection modules cover PII detection, secrets detection, prompt injection patterns, toxic content, PCI-DSS card data, and HIPAA-covered identifiers.

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship out of the box. A preset activates the relevant inspection modules and sets default decisions for common violation types. Teams can override any default to tighten or adjust for their specific environment.

For actions that should not be automatically blocked but also should not run without review, the require-approval decision holds the action in a queue until a human approves or denies it. The agent waits. Nothing executes until the decision is made.

Performance at production scale

A policy enforcement layer is only viable if it does not add meaningful latency to agent workflows. Vaikora runs at P50 of 8ms and P99 of 45ms per enforcement decision, and handles more than 10,000 actions per second. At those numbers, the gate is invisible to the agent and to end users.

The gateway supports 12 LLM providers with no SDK migration. Teams connect Vaikora to an existing agent by routing calls through the proxy and dropping a policy file in place.

What to check in your own environment

If you are running AI agents in production, three questions are worth answering before anything else:

If any of those answers is "no" or "not sure," the Vaikora open-source proxy is a reasonable starting point. Install it from PyPI as `vaikora-guard-mcp`, configure a policy file, and route your first agent through it. The enforcement behavior is visible immediately, and the audit log starts writing from the first decision.

See Vaikora enforce policy on your AI

Open-core AI runtime control. Self-host the MIT gateway free, or run the hosted Control Plane.

Get a demo Self-host the gateway

More from the Vaikora blog