agentmatrix-core 0.7.0.25

Core execution engine for building AI agent applications — MicroAgent, AgentShell protocol, Cerebellum, and skill system

agentmatrix-core

Core execution engine for building AI agent applications.

Let LLMs think. Don't make them write JSON.

AgentMatrix separates reasoning from formatting. The large model thinks in natural language. A smaller model (Cerebellum) translates intent into executable parameters. Two models, each doing what they're best at.

Install

pip install agentmatrix-core

Requires Python 3.12+.

Architecture

┌─────────────────────────────────────────────┐
│  App Layer     Your Application             │
├─────────────────────────────────────────────┤
│  Shell Layer   AgentShell Protocol           │
│                (interface you implement)     │
├─────────────────────────────────────────────┤
│  Core Layer    MicroAgent Engine             │
│                (this package)                │
└─────────────────────────────────────────────┘

• Core Layer — MicroAgent is the execution engine. Pure reasoning loop: think, detect actions, execute, repeat. No I/O, no UI.
• Shell Layer — AgentShell is the protocol you implement to connect Core to the outside world (LLM clients, prompt templates, session storage, etc.).
• App Layer — Your application that wires everything together.

This separation means the same core agent behavior runs anywhere — desktop, terminal, or cloud.

Quick Start

1. Implement AgentShell

AgentShell is the interface between the Core engine and your application:

from agentmatrix.core.agent_shell import AgentShell
from agentmatrix.core.micro_agent import MicroAgent

class MyShell(AgentShell):
    # Implement the required methods:
    # - get_llm_client()    → your LLM backend
    # - get_system_prompt() → prompt template
    # - get_session_store() → session persistence
    # - on_action_result()  → handle action outputs
    # - on_agent_message()  → handle agent responses
    ...

2. Create a MicroAgent and Run

agent = MicroAgent(
    name="my-agent",
    shell=my_shell,
    skills=["file","web-search"],
)

# Start the reasoning loop
await agent.run("List files in the current directory")

3. See a Working Example

A complete terminal agent (~200 lines) is available in the repository:

git clone https://github.com/webdkt/agentmatrix.git
cd tutorial/cli-agent

export OPENAI_API_KEY=sk-xxx
python main.py -m https://endpoint-url:deepseek-v4-pro

Key Modules

Module	Description
core.micro_agent	The execution engine — think, detect actions, execute, repeat
core.agent_shell	Shell protocol — implement this for your app
core.cerebellum	Intent-to-action parameter negotiation
core.action	Action registry and execution
core.session_store	Session persistence interface
core.signals	Event-driven communication (pause, resume, stop)

Key Features

Natural Language Reasoning

The agent's "Brain" reasons entirely in natural language. No JSON output required, no format constraints. A separate "Cerebellum" translates intent into executable parameters.

Pause, Resume, Stop

Any running agent can be paused, resumed, or stopped via signals. State is preserved at safe checkpoints.

Context Auto-Compression

When conversation history grows too large, the system automatically compresses it into "Working Notes" — a dynamic state snapshot generated by the LLM. Tasks can run for hours; the context window never overflows.

Action System

Actions are detected from natural language output via <action_script> blocks. The Cerebellum negotiates parameters with the Brain, handles ambiguity, and executes.

Skill System

Built-in Python skill mixins:

• base — Date/time utilities
• file — File read/write, search
• shell — Shell command execution

Extend with custom Python skills or Markdown-based procedural knowledge.

Dependencies

• pyyaml>=6.0
• python-dotenv>=1.0.0
• requests>=2.31.0
• aiohttp>=3.8.0

Links

• Repository: https://github.com/webdkt/agentmatrix
• Tutorial: tutorial/cli-agent/
• Documentation: docs/
• License: Apache License 2.0