Architecture

Built as a Control Plane. Not Middleware.

A production-grade governance control plane designed for sub-10ms enforcement, horizontal scaling, and fail-closed operation across every cloud.

Request a Demo View Product

High-Level Architecture

Four components. One enforcement point. No exceptions.

Least Trusted
Most Trusted
Untrusted Zone

AI Agent

LangChain · CrewAI · LlamaIndex · Bedrock Agents · Azure AI · GCP ADK · AutoGPT

↓
tool call intercepted
Your Environment · VPC / On-Prem

Execution Gateway

Stateless enforcement proxy deployed in your infrastructure. No data leaves your environment.

Policy Engine
OPA / Rego
Kill Switch
Agent / Session / Fleet
Circuit Breaker
Auto-remediation
<10ms p99 Fail-Closed Stateless No LLM Deterministic
↓
encrypted · mTLS
SaaS · Self-Hosted · Air-Gapped

GovernorAI Control Plane

Centralized governance services. Policy management, compliance automation, audit trail, and fleet orchestration.

Policy Store
Versioned · Git-sync
Event Ingest
Merkle-sealed
Compliance
SOC 2 · GDPR · HIPAA
Approvals
Human-in-the-loop
Dashboards
Fleet observability
Audit Trail
Immutable · Exportable

Core Components

Execution Gateway

Stateless, horizontally scalable enforcement point. Every tool call from every agent passes through the gateway before reaching any target system. <10ms p99 at 1000+ RPS.

  • Fail-closed: no policy match = denied
  • Allow / Deny / Require Approval
  • Horizontal auto-scaling

Policy Engine (OPA/Rego)

Deterministic policy evaluation with zero LLM involvement. YAML policies for common patterns, OPA/Rego for complex logic. 4 evaluation modes.

  • YAML + OPA/Rego hybrid
  • <3ms median evaluation
  • No LLM in the governance loop

Event Ingest (Append-Only)

Immutable, append-only audit trail. Every action, every decision, every outcome logged as structured JSON events. No raw prompts stored.

  • Structured JSON events
  • Merkle-sealed for tamper resistance
  • SOC 2, GDPR, HIPAA-aligned

Control Plane

Centralized management for policies, sessions, kill switches, and approval workflows. API-first design for integration with existing tooling.

  • Policy CRUD with version control
  • Session lifecycle management
  • Kill switch propagation <100ms

Policy Evaluation Flow

Every request follows the same deterministic path.

Request Lifecycle

  1. Intercept — Agent tool call arrives at the Execution Gateway
  2. Identify — Gateway resolves agent ID, session, and policy binding
  3. Evaluate — Policy Engine runs YAML rules + OPA/Rego evaluation
  4. Decide — Allow, Deny, or Require Approval
  5. Log — Structured event emitted to append-only audit trail
  6. Execute or Block — Allowed actions proceed; denied actions return error to agent
JSON 
// 1. Agent sends tool call
{
  "agent_id": "finance-agent-v1",
  "session_id": "sess-123",
  "tool": "erp.process_payment",
  "args": {"amount": 7500, "currency": "USD"}
}

// 2. Gateway evaluates & responds
{
  "action_id": "act-789xyz",
  "decision": "require_approval",
  "reason": "Amount exceeds $5,000 threshold",
  "policy_id": "finance-agent-policy",
  "policy_version": 3,
  "latency_ms": 4
}

Multi-Cloud Transpilation Pipeline

One policy. Every cloud. Semantic parity guaranteed.

1

Author

Write governance policy in YAML or Rego. One canonical format.

2

Compile

GovernorAI compiles to a Canonical Policy Model (CPM).

3

Transpile

CPM transpiled into native artifacts for each target cloud.

4

Deploy

Native enforcement deployed to AWS, Azure, GCP, Databricks, K8s.

5

Reconcile

Continuous drift detection and automatic reconciliation across all clouds.

∞

Monitor

Unified observability across every enforcement point. One dashboard.

Multi-Cloud Deep-Dive →

Deployment Topologies

GovernorAI runs wherever your agents run.

Single-Cloud

Deploy GovernorAI alongside your agents on a single cloud provider. Lowest latency, simplest setup.

Multi-Cloud

Unified governance across AWS, Azure, GCP, Databricks, and Kubernetes. One policy, every cloud.

Self-Hosted / Air-Gapped

Run GovernorAI entirely within your VPC or on-prem infrastructure. No external dependencies. Air-gapped support available.

Performance Architecture

Designed as infrastructure, not middleware. Performance is a feature.

<10ms
p99 latency
<3ms
median eval
1000+
requests/sec
<100ms
kill switch

Stateless Design

Execution Gateway holds no state. Scale horizontally by adding instances. No coordination overhead.

Connection Pooling

Efficient connection management to Policy Store and Event Ingest. Zero-allocation hot paths.

Policy Caching

Hot policies cached in-memory with TTL-based invalidation. Sub-millisecond policy lookups.

See It in Action

Request a demo to see GovernorAI's architecture deployed against your use case.

Request a Demo View Product