Skip to main content

​
Architecture

OmniCoreAgent is built with a modular, extensible architecture designed for production-grade reliability and scalability. It follows a “layered” approach where each component can be swapped or extended without impacting the others.

​
High-Level Overview

OmniCoreAgent acts as the central orchestrator between users, Large Language Models (LLMs), and external tools (via MCP or local functions).

​
Core Layers

​
1. The Reasoning Layer (ReAct)

The heart of OmniCoreAgent is its implementation of the ReAct (Reason + Act) pattern. Instead of a single completion, the agent engages in a loop:
  • Think: Analyze context and user intent.
  • Act: Call tools or sub-agents to gather info or perform actions.
  • Observe: Read tool outputs and update internal state.
  • Answer: Provide the terminal response once the goal is achieved.

​
2. The Storage Layer (Memory & Events)

OmniCoreAgent decouples data persistence from reasoning logic.
  • Memory Router: Manages conversation history. It supports multiple backends (Redis, PostgreSQL, etc.) and can switch between them at runtime.
  • Event Router: Handles real-time telemetry. Every thought and action is emitted as a structured event.

​
3. The Connectivity Layer (MCP & Local)

The framework provides a unified interface for all external actions:
  • MCP Transport: Bridges the gap to the Model Context Protocol ecosystem via stdio, SSE, or Streamable HTTP.
  • Tool Registry: A decorator-driven system for registering Python functions.
  • BM25 Retrieval: A lexical search layer that filters thousands of tools down to a manageable context for the LLM.

​
Data Flow: Processing a Request

1

Initialization

The agent loads its system instructions and tool indexes.
2

User Input

A run(query, session_id) call triggers the workflow.
3

Context Retrieval

The Memory Router fetches previous messages for the session.
4

Tool Discovery

If advanced tool use is enabled, BM25 finds the top relevant tools for the current query.
5

LLM Interaction

The reasoning loop begins. LiteLLM routes prompts to the configured provider.
6

Execution

If the LLM requests a tool, the Tool Orchestrator executes it (local or remote).
7

Event Emission

Throughout the process, the Event Router streams progress updates.
8

Finalization

The agent returns the response and session ID to the application.

​
Enterprise Features

FeatureArchitectural Implementation
Multi-TenancySession IDs isolate conversation states across different users.
ObservabilityNative Opik tracing and Redis Stream event logging.
Fault ToleranceBuilt-in retries for tool calls and automatic connection recovery for MCP.
ScalabilityDistributed memory backends allow agents to scale horizontally across servers.
This modular design allows you to use OmniCoreAgent as a lightweight local library or as the backbone of a high-scale, distributed AI platform.