Skip to content

MAF-INTEGRATION-PROPOSAL.md

Latest commit

 

History

History
106 lines (78 loc) · 4.5 KB

MAF-INTEGRATION-PROPOSAL.md

File metadata and controls

106 lines (78 loc) · 4.5 KB

Microsoft Agent Framework — Governance Middleware Integration

Submission: microsoft/agent-framework#4440 Status: ✅ Implemented (Level 1) — Middleware adapter with 18 passing tests. Awaiting MAF team for Level 2/3. Type: Feature proposal (3-level integration) Sponsor: Patrick Chanezon (VP DevRel) — "+@ShawnHenry to integrate with MAF before you ship this." Date Submitted: March 3, 2026

Context

As part of the OSS approval for microsoft/agent-governance-toolkit, VP Developer Relations Patrick Chanezon specifically requested MAF integration before public release:

"+@Shawn Henry to integrate with MAF before you ship this."

This document outlines the existing integration and proposed deeper collaboration.

What We Built

Three middleware classes that plug directly into Agent Framework's middleware pipeline:

GovernancePolicyMiddleware

  • Hook: process() — intercepts before agent execution
  • Capabilities: Token limits, rate limiting, blocked patterns (regex/glob), content safety
  • OWASP: Mitigates ASI-01 (Agent Hijacking), ASI-02 (Excessive Capabilities)

CapabilityGuardMiddleware

  • Hook: process() — blocks unauthorized tool calls
  • Capabilities: Tool-level allow/deny lists, capability sandboxing
  • OWASP: Mitigates ASI-02 (Excessive Capabilities), ASI-06 (Confused Deputy)

OutputValidationMiddleware

  • Hook: process() — validates agent responses
  • Capabilities: PII detection, output length limits, blocked output patterns
  • OWASP: Mitigates ASI-05 (Insecure Output)

MAFKernel Convenience Class

Configures all three middleware with a single policy object:

from agent_os.integrations.maf_adapter import maf_govern

kernel = maf_govern(agent, policy="policy.yaml")
result = await kernel.invoke(task)

Test Coverage

18 tests passing covering:

  • Policy enforcement (token limits, rate limiting, blocked patterns)
  • Capability guards (tool allow/deny, unknown tool blocking)
  • Output validation (PII detection, length limits)
  • Goal drift detection
  • Middleware chaining and error handling

Integration Proposal

Level 1: Documentation (Immediate)

  • Add Agent Governance Toolkit to Agent Framework's ecosystem/integrations docs
  • Show how governance middleware slots into existing middleware pipeline
  • Effort: 1-2 days
  • Owner: Agent Governance Toolkit team

Level 2: Sample (Short-term)

  • Contribute a governance sample to Agent-Framework-Samples
  • End-to-end example: policy YAML → middleware → agent with audit trail
  • Effort: 1 week
  • Owner: Joint (Agent Governance + MAF team)

Level 3: Native Support (Medium-term)

  • Governance middleware as an optional package (agent-framework[governance])
  • Policy schema validation integrated with Agent Framework's config system
  • Built-in OWASP Agentic Top 10 compliance checks
  • Effort: 2-4 weeks
  • Owner: Joint (requires MAF team architecture guidance)

Why This Matters

  1. Enterprise readiness — Customers need governance before deploying agents in production. This makes MAF enterprise-ready out of the box.
  2. OWASP coverage — The middleware covers 10/10 OWASP Agentic Top 10 risks at the framework level.
  3. VP-approved OSS — This project has been approved for open-source release under microsoft/ org by VP Developer Relations.
  4. Standards alignment — Active proposals at AAIF, LF AI, CoSAI WS4, and OWASP for governance standards.

Existing Code

File Location
MAF Adapter packages/agent-os/src/agent_os/integrations/maf_adapter.py
MAF Tests packages/agent-os/tests/test_maf_adapter.py (18/18 passing)
README MAF listed as primary integration in toolkit README

Related Proposals

Stakeholder Action Link
Marco Casalaina Foundry governance integration Pending discussion
Chris DiBona / Caitie AAIF project candidacy docs/AAIF-PROPOSAL.md
Shawn Henry MAF integration (this document) microsoft/agent-framework#4440

Next Steps

  1. Await MAF team response on preferred integration approach
  2. Submit governance sample to Agent-Framework-Samples
  3. Coordinate with Shawn Henry per VP's direction
  4. Iterate on middleware API based on MAF team feedback