primitives

Agent Primitives

Part of Agent OS - Kernel-level governance for AI agents

Layer 1 Primitive - Shared data models for the Agent OS stack.

Purpose

This package provides foundational data models used across multiple Agent OS components. By extracting these primitives into a dedicated Layer 1 package, we ensure proper dependency layering:

Layer 1 (Primitives): cmvk, emk, caas, agent-primitives
Layer 2 (Infrastructure): iatp, amb, atr
Layer 3 (Kernel): agent-control-plane  
Layer 4/5 (Intelligence): scak, mute-agent

Installation

pip install agent-primitives

Models

Failure Models

Core failure tracking primitives used by iatp, scak, and other components:

from agent_primitives import (
    FailureType,
    FailureSeverity,
    AgentFailure,
    FailureTrace,
)

# Create a failure record
failure = AgentFailure(
    agent_id="agent-123",
    failure_type=FailureType.TIMEOUT,
    error_message="Request timed out after 30s",
    severity=FailureSeverity.MEDIUM,
)

Available Types

• FailureType: Enumeration of failure categories (TIMEOUT, INVALID_ACTION, RESOURCE_EXHAUSTED, etc.)
• FailureSeverity: Severity levels (LOW, MEDIUM, HIGH, CRITICAL)
• AgentFailure: Core failure record with agent ID, type, message, and context
• FailureTrace: Detailed trace including reasoning chain and failed action

Design Principles

1. Zero Agent OS Dependencies: This package only depends on pydantic
2. Backward Compatible: Other packages can re-export these models
3. Type Safe: Full typing support with py.typed marker

License

MIT