Architecture

Neutral. Cross-stack. Deterministic.

No cloud lock-in. No external services on the decision path. No LLM judging another LLM. Verun runs wherever your agents run — and authorizes their actions inline, with negligible overhead.

System Design

One gateway. A three-layer engine.

Decision Path (inline)
Agent Runtime — any framework, any model
LangGraphCrewAIAutoGenOpenAIAnthropicLocal models
↓ proposes a consequential action ↓
Action Control Gateway

Verun Proxy / Sidecar

inline · deterministic · no LLM

Policy Can it?

Tools, thresholds, default-deny

Behavior Does it usually?

Per-agent baseline · anomalies

Mission Should it?

Off-topic tripwire

Decision

ALLOW / HOLD / BLOCK / ESCALATE

Enforcement mode

observe → hold_only → full

Signed receipt

Tamper-evident, async write

OTEL span

Additive, non-blocking

↓ on ALLOW ↓
Tool / API / Model

Receives the original request. Never sees blocked actions.

↓ on HOLD / ESCALATE ↓
Human Review Queue

Approve once · or promote to a standing rule

↓ async, non-blocking ↓
The Data Moat — compounds over time

Per-Agent Behavioral Baseline Store

Every action sharpens each agent's baseline. The more agents flow through Verun, the better the anomaly detection.

Control Plane
Admin Console

React + TypeScript

DashboardReceipt ExplorerHuman Review QueuePolicy ManagerBehavioral BaselinesMission ManagerEvidence Packs
↓ HTTP/REST ↓
Backend API

FastAPI (Python)

JWT AuthScoped API Keys (read/review/admin)Governance Evidence PacksAudit Log
↓ SQL ↓
Storage

SQLite (local) / PostgreSQL (production)

receipts

Signed action decisions

baselines

Per-agent behavioral profiles

missions

Declared agent mandates

policies

Configured policy rules

Design Decisions

Every choice made with intent.

No LLM on the decision path

Using an AI to judge another AI's actions reintroduces bias, non-determinism, latency, and bypassability. Verun's decisions are deterministic — the same input always yields the same decision and the same stated reason. An LLM may help a human understand a decision; it never makes one.

Air-gapped capable

The decision path makes zero external calls, so Verun runs in fully isolated environments. Financial institutions and defense contractors can deploy on secure networks with no external egress required.

Three integration modes

Proxy (zero code changes), Sidecar (K8s network-layer interception via Envoy/Istio ext_authz), and SDK (explicit instrumentation). The proxy model means every agent framework works — they all speak HTTP.

Enforcement mode graduation

Shadow mode lets tenants observe Verun's decisions before any blocking occurs. Graduate from observe → hold_only → full at your own pace. No team deploys blocking infrastructure cold.

Policy mining from observations

After observing real traffic, Verun mines the recorded decisions into a draft policy — an allowlist of the actions your agents actually use, sensible thresholds from what it saw, and default-deny for the rest. Deterministic and explainable, every rule backed by evidence, no LLM. You review and approve.

Behavioral + mission engine

Beyond policy, a per-agent baseline scores each action for behavioral anomalies (velocity, first-seen targets, unusual sequences) while a semantic mission anchor flags drift off the declared purpose. Both are deterministic and both hold rather than silently block.

Receipts as a first-class output

Every decision is a signed, tamper-evident, replayable record — which layer decided and why, under which policy version. Receipts export as Governance Evidence Packs and emit OpenTelemetry spans to any OTLP backend.

Defensibility

The moat is execution and data.

Verun's advantage isn't a single feature — it's the combination. A deterministic engine that's explainable and auditable, behavioral baselines that compound with every decision, and a neutral position across every model provider.

All three are live today and get stronger with usage.

Deterministic engine

Live

No LLM on the decision path. Same input, same verdict, same reason — the only stance defensible to a regulator.

Compounding decision data

Live

Every allow, hold, block, and human ruling sharpens the per-agent baselines. The advantage grows with usage.

Neutral & cross-stack

Live

Governs agents on OpenAI, Anthropic, or local models — a position single-cloud vendors can't hold neutrally.

Integrations

Works with your existing stack.

The proxy model means any agent framework works — they all speak HTTP. One environment variable points agents at the Verun gateway.

LangGraph

Proxy + SDK

CrewAI

Proxy

AutoGen

Proxy

OpenAI / Anthropic

Proxy

MCP clients

MCP gateway

n8n / Make / Zapier

No-code node

Envoy / Istio

Sidecar ext_authz

Any HTTP agent

One-line call

OpenTelemetry

Receipt export

The gateway intercepts agent HTTP traffic transparently. No framework-specific connectors required.

What Verun Sees

We check the action, not the content.

Two questions every buyer asks: what does Verun actually receive, and does our data leave our environment? Straight answers.

Action metadata — not your content

Verun evaluates the action: agent id, action type, target, and parameters (amount, recipient, resource). It does not need your conversations, documents, or customer PII to make a decision. Sensitive fields can be hashed or redacted before they ever reach the engine.

agent_idaction_typetargetparametershash / redact

Runs inside your boundary

If data leaving is a red line, you don't send it anywhere. Verun deploys as a self-hosted container — in your VPC or fully air-gapped. Nothing crosses the line, and receipts stay with you under your own controls.

Self-hosted containerIn-VPCAir-gappedReceipts stay with you

On the honest cert story: we're not SOC 2 or ISO certified yet(both are on the roadmap). If data leaving is a red line, that's moot — run Verun in your own environment and your data stays under your controls.

Built to satisfy your auditors.

Verun's controls and evidence map to the governance frameworks enterprises are measured against — see the alignment.