Skip to content

INDEPENDENCE.md

Latest commit

 

History

History
84 lines (62 loc) · 3.91 KB

INDEPENDENCE.md

File metadata and controls

84 lines (62 loc) · 3.91 KB

Independence & Dependency Policy

The Agent Governance Toolkit is designed to be a standalone governance standard with zero vendor lock-in in its core packages.

Core Independence Rule

No core package may import from any vendor framework at module level without a try/except guard.

Core paths (agent_os/, agentmesh/, agent_hypervisor/, agent_sre/) must function with only standard-library and widely-adopted infrastructure dependencies (pydantic, cryptography, pyyaml).

Package Independence Matrix

Package Hard Vendor Deps Status
agent-os-kernel (Python) None — pydantic only ✅ Independent
agentmesh-platform (Python) None — pydantic + cryptography ✅ Independent
agent-hypervisor (Python) None — pydantic only ✅ Independent
agent-sre (Python) None — pydantic + structlog ✅ Independent
agent-governance-toolkit (Python) None — pydantic only ✅ Independent
agentmesh (Rust) None — pure crypto + serde ✅ Independent
agentmesh-mcp (Rust) None — pure crypto + serde ✅ Independent
agentmesh (Go) None — yaml.v3 only ✅ Independent
@agentmesh/sdk (TypeScript) None — zero runtime deps ✅ Independent
Microsoft.AgentGovernance (.NET) None — YamlDotNet only ✅ Independent

Adapter Pattern

Framework integrations are published as separate packages that depend on AGT core + the target framework. This keeps the core clean while enabling any ecosystem:

Adapter Package Framework Install
langchain-agentmesh LangChain pip install langchain-agentmesh
llamaindex-agentmesh LlamaIndex pip install llamaindex-agentmesh
crewai-agentmesh CrewAI pip install crewai-agentmesh
openai-agents-agentmesh OpenAI Agents pip install openai-agents-agentmesh
pydantic-ai-governance Pydantic AI pip install pydantic-ai-governance
aps-agentmesh APS pip install aps-agentmesh
scopeblind-protect-mcp ScopeBlind pip install scopeblind-protect-mcp

Adapters must use try/except for all framework imports so they fail gracefully when the framework isn't installed.

Observability Integrations

agent-sre supports 12+ observability platforms as optional dependencies. None are required:

pip install agent-sre                    # Core only — zero vendor deps
pip install agent-sre[arize]             # + Arize Phoenix
pip install agent-sre[langfuse]          # + Langfuse
pip install agent-sre[wandb]             # + Weights & Biases
pip install agent-sre[datadog]           # + DataDog
pip install agent-sre[full]              # All integrations

All vendor integrations live under agent_sre/integrations/ and use try/except import guards.

Policy Engine Backends

The policy engine supports multiple backends without hard dependencies:

Backend Dependency Required?
Native YAML Built-in ✅ Always available
OPA/Rego opa CLI (external) Optional
Cedar cedarpy Optional

What This Means for Adopters

  1. pip install agent-governance-toolkit gives you full governance with zero vendor deps
  2. Add framework adapters only for frameworks you actually use
  3. Core packages will never require LangChain, OpenAI, Anthropic, or any specific LLM provider
  4. Rust, Go, .NET, and TypeScript SDKs follow the same zero-vendor-dep principle

Contributing

When adding new code to core packages:

  • ❌ Do not add vendor framework imports at module level
  • ✅ Use try/except guards for any optional imports
  • ✅ Place framework-specific code in integrations/ directories or separate adapter packages
  • ✅ Add new dependencies to [project.optional-dependencies], not [project.dependencies]