MS Agent Governance Toolkit vs Google AX — two opposite layers of 2026 agent infrastructure

notion image

MS Agent Governance Toolkit vs Google AX — two opposite layers of 2026 agent infrastructure

Two of the largest agent infrastructure projects of 2026 dropped within weeks
of each other. They share the keyword "agent runtime" but sit on opposite
sides of the stack. Reading them as competitors misses the point — they're
the two halves of the same architecture.

TL;DR

  • MS Agent Governance Toolkit (AGT) — a policy & security plane that decides what an agent is allowed to do.
  • Google AX (Agent eXecutor) — a distributed execution plane that runs the agent and survives crashes.
  • Not competitors. AGT explicitly ships a Google ADK integration; AX leaves governance out of scope. The intended deployment is AGT on top of AX, not one or the other.

1. Positioning

Dimension
MS Agent Governance Toolkit
Google AX
Layer
Governance / security side-car
Execution runtime (orchestration core)
Core value
Action control, compliance
Reliability, automatic recovery
Announced
2026-04-02 (MS Open Source blog)
KubeCon NA 2025-11 → expanded 2026
License
MIT
Apache 2.0
Maturity
Public Preview, v3.7.0 (May 2026)
Active early development, pre-v1.0 (breaking changes expected)
Explicit non-goals
Prompt guardrails, content safety
Policy/identity beyond audit call wiring
The split is intentional. AGT's documentation states that prompt-based safety
("just tell the model to follow the rules") yields a 26.67% policy violation
rate, whereas application-layer enforcement reaches 0.00%. AX takes the
inverse stance: it doesn't try to police what the agent says, it makes sure the
agent's process survives and its state is recoverable.

2. Microsoft Agent Governance Toolkit

Features

Capability
Detail
Policy Engine
YAML, OPA/Rego, Cedar. Single rule eval at 0.012 ms (72K ops/sec), full enforce at 0.091 ms, 50-agent parallel run at 35,481 ops/sec
Zero-Trust Identity
Ed25519 plus post-quantum ML-DSA-65, trust score 0–1000
Execution Sandboxing
4-tier privilege rings, Saga orchestration, kill-switch
Agent SRE
SLO / error budget / chaos engineering / circuit breakers
MCP Security Scanner
Tool poisoning, lookalike-name attacks, hidden-instruction detection
Shadow AI Discovery
Auto-detection of unregistered agents
Compliance coverage
OWASP Agentic Top 10 ASI-01 ~ ASI-10 in full (first toolkit to do so, validated by 13K+ tests)

Tech stack — polyglot across 5 languages

Language
Package
Install
Python
agent-governance-toolkit[full]
pip install ...
TypeScript
@microsoft/agent-governance-sdk
npm install ...
.NET
Microsoft.AgentGovernance
dotnet add package ...
Rust
agent-governance
cargo add ...
Go
agent-governance-toolkit
go get ...
Copilot CLI
@microsoft/agent-governance-copilot-cli
npx ...
Claude Code
dedicated plugin

Python package layout

  • Agent OS — policy engine, capability model, MCP security gateway
  • AgentMesh — zero-trust identity, trust scoring, mesh coordination
  • Agent Runtime — privilege rings, saga orchestration, execution control
  • Agent SRE — SLOs, error budgets, chaos, circuit breakers
  • Agent Compliance — OWASP verification, policy linting, integrity checks

Framework integrations

Each adapter hooks into the framework's native extension point — no agent
code rewrite required:
  • Microsoft Agent Framework, Semantic Kernel
  • LangChain / LangGraph (callback handlers)
  • CrewAI (task decorators)
  • AutoGen, OpenAI Agents SDK
  • Google ADK (plugin system), LlamaIndex, Haystack, Mastra, Dify
  • Azure AI Foundry, AWS Bedrock

Strengths

  • Full coverage of OWASP Agentic Top 10 → NIST AI RMF and EU AI Act compliance ready out-of-the-box
  • 5 languages × almost every major agent framework — drops into existing code
  • Sub-millisecond enforcement, Microsoft-signed builds, MIT
  • Only option that ships governance plugins for Claude Code and Copilot CLI

Weaknesses

  • Provides no execution plane — a separate runtime is still required
  • Content safety / prompt injection is explicitly out of scope
  • Still Public Preview at v3.7.0 — breaking changes possible before GA
  • Maintaining 5 language implementations risks feature drift across them

3. Google AX (Agent eXecutor)

Features

  • Distributed runtime — Controller, Skill, Tool, and Agent each run as independent processes
  • Single-writer architecture — guarantees state consistency
  • Event log on SQLite — durable record of every execution, enables automatic resume after interruption
  • gRPC actor communication — uniform interface for local and remote agents (default localhost:8494)
  • A2A protocol bridge — interoperates with any A2A-compliant external agent
  • Native on GKE Agent Sandbox / Agent Substrate — 16× growth in 5 months of preview

Tech stack — Go-centric monolithic core

  • Language: Go 84.8%, Python 12.6%
  • Transport: gRPC (proto/ax.proto)
  • Storage: SQLite (event log)
  • AI: Gemini API (built-in planner)
  • Deployment: Kubernetes / Agent Substrate on GKE (recommended)
  • Install: go install github.com/google/ax/cmd/ax@latest

Components

  • Controller — coordinates execution, manages state
  • EventLog — SQLite-backed durable event store
  • AgentService — gRPC interface for remote agents
  • Planner — default Gemini-backed planning agent

Supported SDKs

Native AX agent, Google ADK (Python), A2A-compliant external agents,
Google Colab (experimental).

Strengths

  • Long-running execution and failure recovery as first-class citizens — event log handles automatic resume
  • gRPC + Kubernetes-native — horizontal scale and multi-node distribution feel natural
  • A2A bridge for cross-ecosystem interoperability
  • Combined with GKE Agent Sandbox, sandboxing and scaling are delegated to the substrate
  • Single Go core keeps operational surface simple

Weaknesses

  • Pre-v1.0 with major breaking changes announced upfront
  • Go toolchain and Kubernetes are effectively required (Substrate is the recommended path)
  • Policy / governance / identity is limited to "wiring audit calls" — no real enforcement
  • Optimized around Google ADK — integrations for other frameworks rely on the A2A bridge
  • Documentation and examples are still sparse (early development)

4. Decision guide

Situation
Pick
Already running LangGraph / CrewAI / Semantic Kernel agents and worrying about compliance, audit, MCP risk
AGT (one-line install)
Running dozens to hundreds of agents on GKE, watching them die and losing state
AX (event-log resume)
Enterprise full stack — both governance and reliable execution
AGT + AX together (AGT explicitly lists ADK integration → same-stack deployment is the intended path)
Need to attach governance to Claude Code or Copilot CLI
AGT — the only option with dedicated plugins

5. Why this matters in 2026

  • Two hyperscalers open-sourced opposite abstraction layers at almost the same moment. That's a signal that 2026 agent infrastructure is splitting cleanly into runtime (AX) and governance (AGT).
  • AGT explicitly shipping a Google ADK integration is the strongest hint that the two were designed to compose, not compete.
  • OWASP's December 2025 release of the Agentic Top 10 is becoming the de-facto standard — AGT is its first comprehensive implementation.
  • AX's 16× growth on GKE Agent Sandbox confirms the trend of "agents as container workloads", which makes a Kubernetes-native execution plane the new baseline.

Sources

TECH LINKS