ABrain Documentation

Welcome to the ABrain documentation. This documentation now focuses on the current foundations stack in this repository: canonical agent model, Flowise interop, decision layer, execution layer and learning system around the hardened core path. Current foundations release: v1.1.0.

Overview

ABrain is a multi-agent orchestration system that combines typed tool execution, deterministic policy checks and neural ranking to provide a controlled execution framework. The current repository emphasizes:

• Canonical AgentDescriptor modeling
• Deterministic planning and candidate filtering
• Always-on NeuralPolicyModel ranking
• Static execution adapters behind a separated execution layer
• Feedback-driven learning with best-effort training
• A hardened dispatcher and tool layer
• Security-focused adapter integration and interface boundaries

Key Features

Foundations Release

• Canonical agent model in core/decision/*
• Flowise import/export as a thin interoperability layer
• Planner, CandidateFilter, NeuralPolicyModel and RoutingEngine
• Execution engine with static adapter registry
• Learning dataset, reward model, online updater and trainer
• Hardened core dispatch path via services/core.py

Getting Started

For the current developer-facing foundations path, start with Project Overview, Development Setup and Foundations Release Notes.

Architecture Overview

The current foundations pipeline consists of these key stages:

graph TD
    A[Task] --> B[Planner]
    B --> C[CandidateFilter]
    C --> D[NeuralPolicyModel]
    D --> E[RoutingDecision]
    E --> F[ExecutionAdapter]
    F --> G[ExecutionResult]
    G --> H[Feedback Loop]
    H --> I[Training Dataset]

For more details about the current system architecture, see the Project Overview.

Contributing

We welcome contributions! Please see our Contributing Guide for details on how to get involved.

License

This project is licensed under the MIT License - see the local LICENSE file for details.