superagentx-policy-engine 1.0.3

SuperAgentX Policy Engine. Enterprise-grade AI Agent Governance Framework. Govern AI agents with policies, approvals, threat detection, cost controls, memory protection, PII, compliance protections and multi-agent security.

SuperAgentX Policy Engine

AI Governance Framework for Autonomous AI Systems

SuperAgentX Policy Engine is an AI Agent Governance Framework designed to secure, control, and govern autonomous AI systems.

Unlike traditional authorization frameworks, SuperAgentX Policy Engine understands:

• Agents
• Workflows
• LLMs
• Tools
• Memory
• Multi-Agent Systems
• Human Approvals
• AI Threats

---

Why This Exists

Modern AI Agents can:

• Execute shell commands
• Access databases
• Call APIs
• Modify memory
• Spawn new agents
• Communicate with other agents
• Spend money through LLM usage
• Access enterprise systems

Traditional authorization systems were built for users and applications.

They were not built for:

Agent -> Agent
Agent -> Tool
Agent -> LLM
Agent -> Memory
Agent -> Workflow
Agent -> Enterprise Systems

SuperAgentX Policy Engine introduces a governance layer between AI agents and execution.

---

Architecture

Agent
   │
   ▼

Governance Layer

   ├── Threat Detection
   ├── Rate Limiting
   ├── Quota Management
   ├── Policy Evaluation
   ├── Human Approval (HITL)
   ├── Cost Governance
   └── Audit Logging

   │
   ▼

Tool / LLM / Database / API

---

Features

Policy-Based Governance

IAM-style policies.

Human-in-the-Loop (HITL)

Require approvals before risky actions.

Threat Detection

Detect:

• Prompt Injection
• Jailbreak Attempts
• Data Exfiltration
• Credential Theft
• Agent Abuse
• Recursive Agent Attacks

Cost Governance

Control AI spending.

Tool Governance

Control access to tools.

Memory Governance

Protect sensitive memory.

Agent Governance

Control:

• agent:execute
• agent:spawn
• agent:communicate

Audit Trail

Track every decision.

Async First

Built for high-scale agent systems.

---

Installation

pip install superagentx-policy-engine

---

Quick Start

Load Policy

from superagentx_policy_engine.policy_engine import PolicyEngine
from superagentx_policy_engine.store.file_store import FilePolicyStore

engine = PolicyEngine()

policy = FilePolicyStore.load(
    "policies/allow_search.json"
)

await engine.add_policy_document(policy)

---

Policy Format

{
  "version": "2026-01-01",
  "statements": [
    {
      "sid": "AllowWebSearch",
      "effect": "Allow",
      "action": [
        "tool:execute"
      ],
      "resource": [
        "tool:web_search"
      ]
    }
  ]
}

---

Effects

Supported effects:

Allow
Deny
Approve

Priority:

Deny
 >
Approve
 >
Allow

---

Human Approval Example

{
  "sid":"LargeSpendApproval",
  "effect":"Approve",
  "action":["llm:invoke"],
  "resource":["*"],
  "condition":{
    "NumericGreaterThan":{
      "context.custom.estimated_cost":100
    }
  },
  "approval":{
    "role":"finance_manager"
  }
}

---

Tool Governance

Block dangerous tools.

{
  "sid":"BlockShell",
  "effect":"Deny",
  "action":["tool:execute"],
  "resource":["tool:shell"]
}

---

Memory Governance

{
  "sid":"ProtectCustomerMemory",
  "effect":"Deny",
  "action":["memory:delete"],
  "resource":["memory:customer-*"]
}

---

Agent Spawn Protection

{
  "sid":"SpawnProtection",
  "effect":"Deny",
  "condition":{
    "NumericGreaterThan":{
      "context.custom.spawn_count":50
    }
  }
}

---

Threat Detection

Built-in threat protection.

Signature Detection

Detects known attacks.

Rule Detection

Detects suspicious behavior.

LLM Detection

Powered by LiteLLM.

Supported providers:

• OpenAI
• Claude
• Gemini
• Bedrock
• Azure OpenAI
• Ollama
• Groq
• DeepSeek

---

Multi-Stage Threat Detection

detector = EnsembleThreatDetector(
    detectors=[
        SignatureThreatDetector(),
        RuleThreatDetector(),
        LiteLLMThreatDetector(
            model="gemini/gemini-2.5-pro"
        )
    ]
)

Pipeline:

Request
   │
   ▼

Signature Detector
   │
   ▼

Rule Detector
   │
   ▼

LLM Detector
   │
   ▼

Policy Engine

---

Policy Hierarchy

Global Policy
      │
      ▼
Workflow Policy
      │
      ▼
Agent Policy
      │
      ▼
Engine Policy
      │
      ▼
Handler Policy

---

Multi-Agent Governance

Govern agent interactions.

Research Agent
      │
      ▼
Fact Check Agent
      │
      ▼
Summary Agent
      │
      ▼
Report Agent

Policies can control:

agent:execute
agent:spawn
agent:communicate

---

SuperAgentX Integration

AgentXPipe
     │
     ▼

workflow:start
     │
     ▼

agent:execute
     │
     ▼

engine:execute
     │
     ▼

tool:execute

Governance checks can happen at every stage.

---

Use Cases

AI Workforce Governance

Control autonomous teams.

FinOps

Control AI spending.

Production Change Management

Require approvals before production access.

Secure Tool Usage

Restrict dangerous operations.

Multi-Agent Security

Govern agent-to-agent interactions.

Memory Protection

Prevent accidental deletion.

Enterprise AI Governance

Apply policies consistently across all agents.

---

Examples

01_allow
02_deny_override
03_hitl_approval
04_tool_governance
05_cost_governance
06_spawn_protection
07_chain_depth_protection
08_memory_governance
09_threat_detection
10_budget_control
11_prod_database_approval
12_multi_agent_governance
13_showcase_demo
14_llm_threat_detection
15_multi_stage_threat_detection

---

Roadmap

Current

• Allow
• Deny
• Approve
• Conditions
• Threat Detection
• Rate Limiting
• Quotas
• Audit Logs
• LiteLLM Integration

Upcoming

• Policy Registry
• Policy Versioning
• Policy Signing
• Workflow Policies
• Agent Policies
• Engine Policies
• Handler Policies
• ReBAC
• Cedar Compatibility

---

Comparison

Feature	SuperAgentX	OPA	Cedar
Agent Governance	✅	❌	❌
HITL Approvals	✅	❌	❌
Threat Detection	✅	❌	❌
LLM Governance	✅	❌	❌
Cost Governance	✅	❌	❌
Multi-Agent Governance	✅	❌	❌

---

Contributing

Contributions are welcome.

Please open issues, discussions, and pull requests.

---

License

SuperAgentX Policy Engine is released under the MIT License.

---

Built for the next generation of autonomous AI systems.