Modules
Building Blocks for AI Governance
Agent OS is built on a modular architecture where each component solves a specific problem and can be used independently or together. Mix and match based on your needs.
Modular Architecture
Four layers of abstraction—from primitives to applications.
Module Independence Principle
Each module is a standalone package with zero cross-dependencies.
Monolithic Approach
# You want memory, but get everything
pip install mega-ai-framework # 500MB
# Pulls in: ML models, databases,
# message queues, vector stores...Agent OS Approach
# Install only what you need
pip install agent-os-emk # Just memory
pip install agent-os-cmvk # Just verification
# Each module: < 50KBCore Modules
Four specialized modules, each solving a critical AI governance challenge.
CMVK
Cross-Model Verification Kernel
Detect hallucinations and ensure accuracy by comparing outputs across multiple LLMs. Configurable consensus thresholds and model weighting.
from agent_os.cmvk import verify
result = await verify(
prompt="What is 2+2?",
models=["gpt-4", "claude-3"],
threshold=0.8
)
# result.consensus: True
# result.confidence: 0.95EMK
Episodic Memory Kernel
Immutable, auditable agent memory with semantic search and time-travel debugging. Perfect for compliance and reproducibility.
from agent_os.emk import Memory
memory = Memory(agent_id="agent-1")
await memory.store(
content="User requested refund",
metadata={"type": "action"}
)
# Time-travel: replay any momentIATP
Inter-Agent Trust Protocol
Cryptographic identity and message signing for multi-agent systems. Prevent impersonation, verify message integrity, establish trust chains.
from agent_os.iatp import Identity
agent = Identity.create("agent-1")
signed = agent.sign({
"action": "transfer",
"amount": 100
})
# Recipient verifies signatureAMB
Agent Message Bus
Decoupled, async communication between agents. Supports Redis, Kafka, NATS, and SQS backends with unified API.
from agent_os.amb import MessageBus
bus = MessageBus(backend="redis")
await bus.publish(
topic="tasks",
message={"task": "analyze"}
)
# Agents subscribe to topicsInstallation
Install the full kernel or individual modules—your choice.
Full Kernel (All Modules)
# Install everything
pip install agent-os-kernel
# Includes: CMVK, EMK, IATP, AMB + KernelSpaceIndividual Modules
# Cross-Model Verification
pip install agent-os-cmvk
# Episodic Memory
pip install agent-os-emk
# Inter-Agent Trust Protocol
pip install agent-os-iatp
# Agent Message Bus
pip install agent-os-ambWith Optional Backends
# EMK with PostgreSQL support
pip install agent-os-emk[postgres]
# AMB with Kafka support
pip install agent-os-amb[kafka]
# CMVK with all LLM providers
pip install agent-os-cmvk[all]Module Comparison
Choose the right modules for your use case.
| Module | Problem Solved | Use When | Size |
|---|---|---|---|
| CMVK | Hallucination detection | High-stakes decisions, fact verification | ~25KB |
| EMK | Audit trails & debugging | Compliance, debugging, reproducibility | ~30KB |
| IATP | Agent impersonation | Multi-agent systems, zero-trust environments | ~20KB |
| AMB | Agent coordination | Distributed agents, async workflows | ~35KB |
Combining Modules
Modules compose naturally for powerful combinations.
from agent_os import KernelSpace
from agent_os.cmvk import verify
from agent_os.emk import Memory
from agent_os.iatp import Identity
from agent_os.amb import MessageBus
# Initialize modules
kernel = KernelSpace(policy="strict")
memory = Memory(agent_id="finance-agent")
identity = Identity.create("finance-agent")
bus = MessageBus(backend="redis")
@kernel.register
async def finance_agent(task: str):
# Verify critical decisions with CMVK
decision = await verify(
prompt=f"Should we approve: {task}",
models=["gpt-4", "claude-3"],
threshold=0.9
)
# Log to immutable memory
await memory.store(
content=f"Decision: {decision.result}",
metadata={"consensus": decision.confidence}
)
# Sign and broadcast result
signed_msg = identity.sign({"result": decision.result})
await bus.publish("decisions", signed_msg)
return decision.result