Skip to content

LFAI-PROPOSAL.md

Latest commit

 

History

History
107 lines (83 loc) · 5.65 KB

LFAI-PROPOSAL.md

File metadata and controls

107 lines (83 loc) · 5.65 KB

LF AI & Data Foundation — Sandbox Project Proposal

Submission: lfai/proposing-projects#102 Status: Open — awaiting TAC review Requested Level: Sandbox Date Submitted: March 2, 2026

Summary

Agent Governance Toolkit is an open-source governance kernel for autonomous AI agents providing runtime policy enforcement, capability sandboxing, inter-agent trust verification, and kill-switch controls.

The ecosystem consists of 5 interoperating packages:

Package Purpose PyPI
Agent OS Core governance kernel (policy engine, capability sandbox, VFS) agent-os
Agent Mesh Inter-agent trust layer (DID identity, IATP protocol) agent-mesh
Agent Runtime Execution isolation (ring model, kill switch) agentmesh-runtime
Agent SRE Observability & reliability (circuit breakers, anomaly detection) agent-sre
Agent Governance Meta-framework & compliance mapping agent-governance

Key Metrics

  • 82+ GitHub stars, 30+ forks across 5 repos
  • 9,400+ clones in 14 days
  • 5 PyPI packages published
  • MCP server on npm + Glama listing
  • 10/10 OWASP Agentic Top 10 risks covered
  • 4 external contributors
  • All repos: MIT license, CI/CD, branch protection, code of conduct

Alignment with LF AI & Data Mission

As AI agents become increasingly autonomous, governance infrastructure is critical for safe deployment. The Agent Governance Toolkit provides this as a neutral, open-source project — preventing vendor lock-in and enabling a shared governance standard.

Why LF AI & Data?

  1. Neutral governance home — The toolkit needs a vendor-neutral foundation as it moves from Microsoft personal repos to a community project
  2. Cross-framework — Integrations with LangChain, CrewAI, AutoGen, Google ADK, PydanticAI, Mastra, OpenAI Agents SDK, and Microsoft Agent Framework
  3. Standards alignment — Active proposals at AAIF, CoSAI/OASIS WS4, and OWASP
  4. Enterprise readiness — Runtime policy enforcement, audit trails, and compliance mapping are table stakes for enterprise AI agent deployment

Architecture

┌─────────────────────────────────────────────┐
│              Agent Governance                │
│         (Meta-framework + Compliance)        │
├──────────┬──────────┬───────────┬───────────┤
│ Agent OS │  Agent   │  Agent    │  Agent    │
│ (Kernel) │  Mesh    │ Runtime   │  SRE      │
│          │ (Trust)  │(Isolation)│(Observe)  │
├──────────┴──────────┴───────────┴───────────┤
│         Framework Integrations               │
│  MAF · LangChain · CrewAI · ADK · MCP       │
└─────────────────────────────────────────────┘

OWASP Agentic Top 10 Coverage

Risk Description Package
ASI-01 Agent Hijacking Agent OS (PolicyEngine)
ASI-02 Excessive Capabilities Agent OS (CapabilitySandbox)
ASI-03 Insecure Communication Agent Mesh (IATP + DID)
ASI-05 Insecure Output Agent Runtime (OutputValidator)
ASI-06 Confused Deputy Agent OS (CapabilityGuard)
ASI-07 Identity Spoofing Agent Mesh (TrustScorer)
ASI-08 Unbounded Autonomy Agent Runtime (KillSwitch)
ASI-09 Missing Audit Trails Agent SRE (IncidentTimeline)
ASI-10 Cascading Hallucinations Agent SRE (CircuitBreaker)

Test Coverage

Package Tests Coverage
Agent OS 1,327 Policy engine, capability sandbox, VFS
Agent Mesh 476 Trust scoring, DID identity, IATP
Agent Runtime 489 Ring model, kill switch, sagas
Agent SRE 1,071 SLOs, chaos testing, circuit breakers
Agent Governance 537 Compliance mapping, meta-framework
Total 3,900+

Community Traction

  • Framework integration PRs: AutoGen, CrewAI, MetaGPT, OpenAI Swarm
  • Observability integrations: OpenLit (PR #1037), Logfire, HolmesGPT
  • Awesome-list PRs: 15+ across major curated lists
  • Standards submissions: AAIF, LF AI, CoSAI WS4, OWASP ASI
  • Medium articles driving 598+ views per post

Project Governance

  • License: MIT
  • Primary Maintainer: Agent Governance Toolkit Team (Microsoft)
  • Code of Conduct: Contributor Covenant v2.1
  • CI/CD: GitHub Actions on all repos
  • Branch Protection: Required reviews, status checks

Links