Skip to content

fact-sheet.md

Latest commit

 

History

History
129 lines (97 loc) · 7.57 KB

fact-sheet.md

File metadata and controls

129 lines (97 loc) · 7.57 KB

Agent Governance Stack — Fact Sheet

Open-Source Governance for Production AI Agents

One install. Four layers of protection. MIT licensed.

The Stack

Component Purpose Tests Key Metric
Agent OS Governance kernel — policy enforcement, capability security, audit trails 2,000+ 0% policy violations in governed agents
AgentMesh Trust & identity — zero-trust communication, encrypted channels 500+ DID-based cryptographic auth
Agent Runtime Runtime isolation — execution rings, resource limits, kill switches 200+ Process-level isolation
Agent SRE Reliability engineering — SLOs, chaos engineering, incident response 329+ 11 observability platforms
Agent Governance Unified installer — single package for the complete stack — pip install agent-governance-toolkit[full]

Key Capabilities

  • Prompt injection detection — 7 detection strategies (heuristic, ML classifier, canary tokens, perplexity analysis, and more)
  • Policy-as-code — YAML rule sets for capability control, tool allowlists, and output filtering
  • SLO/error budget management — Define availability and latency targets for agents with automated enforcement
  • Chaos engineering + adversarial testing — Inject faults, simulate attacks, and validate governance under stress
  • Circuit breakers and cascade protection — Prevent single-agent failures from cascading across multi-agent systems
  • DID-based cryptographic agent identity — Decentralized identifiers for verifiable, non-repudiable agent authentication
  • OpenTelemetry-native observability — Export governance telemetry as traces, metrics, and logs to any OTEL-compatible backend
  • 12+ framework adapters — LangChain, CrewAI, AutoGen, Semantic Kernel, LlamaIndex, Haystack, OpenAI Agents SDK, Google ADK, MCP, A2A, and more
  • Behavioral anomaly detection — ML-based detection of unusual agent behavior patterns
  • REST API for language-agnostic integration — Use from any language, not just Python

OWASP Agentic AI Top 10 Coverage

10/10 risks fully covered (ASI01–ASI10)

Risk Coverage Mechanism
ASI01 Agent Hijacking ✅ Full Multi-strategy prompt injection detection
ASI02 Privilege Escalation ✅ Full Capability-based access control
ASI03 Insecure Communication ✅ Full mTLS, encrypted channels, trust scoring
ASI04 Insufficient Identity ⚠️ Partial DID-based identity (SPIFFE alternative supported)
ASI05 Unsafe Code Execution ✅ Full Sandboxed execution with resource limits
ASI06 Excessive Autonomy ✅ Full Human-approval policies
ASI07 Data Leakage ✅ Full Output scanning, PII redaction
ASI08 Lack of Observability ✅ Full Audit logging, OpenTelemetry
ASI09 Resource Exhaustion ✅ Full Resource governor, rate limiting
ASI10 Lack of Error Handling ✅ Full Circuit breakers, SLOs, error budgets

See the full OWASP Implementation Guide for code examples.

Differentiators

  1. Only full-stack governance solution — Kernel + trust mesh + runtime isolation + SRE in a single, integrated platform. No other project covers all four layers.

  2. SRE discipline for AI agents — SLOs, error budgets, chaos engineering, and incident response automation applied to agentic AI workloads. This is unique in the market — no other tool brings SRE practices to AI agents.

  3. Zero vendor lock-in — MIT license, open protocols (OpenTelemetry, DID, REST), pluggable architecture. Works with any LLM provider, any framework, any cloud.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    agent-governance                          │
│               pip install agent-governance-toolkit[full]          │
├─────────────────────────────────────────────────────────────┤
│  ┌─────────────────┐          ┌─────────────────────────┐   │
│  │  Agent OS Kernel │◄────────►│  AgentMesh Platform     │   │
│  │  Policy Engine   │          │  Zero-Trust Identity    │   │
│  │  Injection Detect│          │  Mutual TLS             │   │
│  │  Audit Logging   │          │  Trust Scoring           │   │
│  └────────┬─────────┘          └────────────┬────────────┘   │
│           │                                 │                │
│  ┌────────▼─────────┐          ┌────────────▼────────────┐   │
│  │ Agent Runtime     │          │  Agent SRE              │   │
│  │ Execution Rings   │          │  SLO Enforcement        │   │
│  │ Resource Limits   │          │  Chaos Engineering      │   │
│  │ Kill Switch       │          │  Anomaly Detection      │   │
│  └──────────────────┘          └─────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Getting Started

# Install the full stack
pip install agent-governance-toolkit[full]

# Or install individual components
pip install agent-os-kernel          # Governance kernel only
pip install agentmesh-platform       # Trust mesh only
pip install agentmesh-runtime         # Runtime isolation only
pip install agent-sre                # Reliability engineering only
from agent_os import StatelessKernel, ExecutionContext

kernel = StatelessKernel()
ctx = ExecutionContext(agent_id="my-agent", capabilities=["read", "write"])
result = kernel.execute(ctx, action="call_tool", tool="search", args={"q": "query"})

Links

Resource URL
GitHub github.com/microsoft/agent-governance-toolkit
Documentation github.com/microsoft/agent-governance-toolkit/agent-os-docs
PyPI pypi.org/project/ai-agent-governance
OWASP Mapping OWASP Implementation Guide
Architecture Guide Enterprise Reference Architecture
CNCF Proposal CNCF Sandbox Proposal
Author github.com/microsoft/agent-governance-toolkit

Part of the Agent Governance ecosystem — Building the governance layer for the agentic era