OmniAgent

OmniAgent is an open-source self-evolving Agent framework inspired by OpenClaw. It's the only agent that implements full-dimensional self-evolution (OmniEvolve):

• Proactive Memory: A dual-path alignment mechanism based on explicit interactive feedback and implicit LLM induction enables proactive memory and self-evolving
• Skill Self-Evolution: Through automatic creation, inspection, and repair of skills during interaction, skills evolve in real-time
• Context Self-Evolution: Built on a multi-layer information stack architecture, leveraging real-time user interaction feedback and LLM summarization feedback to continuously update memory and user preferences — achieving self-evolving Personalization Context
• BrainModel Self-Evolution: Through a novel online reinforcement learning feedback loop, the BrainModel iterates dynamically during interaction

Together, these enable full-dimensional (Skill, Context, BrainModel) self-evolution of the Agent. Additionally, Hyper Harness and Deep Reflexion modules enhance system safety and task success rate:

• Hyper-Harness: An efficient, safe, and intelligent execution scaffold that provides systematic support for complex tasks
• Deep Reflexion: A dual-layer reflective architecture — real-time risk interception and failure-to-insight conversion — providing a robust guarantee for task success rate

OmniAgent V.S. OpenClaw V.S. Hermes

Core Features

OmniEvolve (Full-Dimensional Self-Evolution): The agent evolves continuously through interaction, and safety hardens dynamically.

• Proactive Memory: Based on a multi-layer information stack, a dual-path alignment mechanism of explicit user feedback and implicit LLM induction enables autonomous precipitation and continuous self-evolution of user profiles and memory
• Skill Self-Evolution: Through pattern extraction from high-frequency action sequences, skills are natively auto-generated; leveraging dual-path feedback from user interaction and LLM diagnosis, skills are automatically diagnosed and repaired
• Personalization Context: Through real-time capture of multi-dimensional preference signals, an adaptive personalized context is constructed, achieving precise alignment between the Agent-Loop and individual user preferences
• BrainModel Self-Evolution: Through a novel online reinforcement learning (GRPO + PRM) feedback loop, the BrainModel achieves closed-loop self-evolution during interactive use

Hyper-Harness (Super Scaffold): A more efficient, safe, and intelligent Harness engine.

• Progressive Context Loading: A design pattern inspired by Anthropic Claude Skills — Progressive Disclosure — loading on demand in graduated stages
• Dynamic Multi-Agent: Introducing Sentinel (planning) and Guardian (safety) agents that dynamically analyze task complexity and risk level, activating in real-time to improve success rate and safety
• Dynamic Concurrent Tool Execution: Auto-resolves inter-tool dependencies, shifting from serial waiting to async parallel invocation, reducing latency for long-chain tasks
• Four-Layer Dynamic Security Scanning: LLM intelligent review → Policy engine → Interactive approval → Execution sandbox. Through trust-level classification, different Skills apply different security policies. Security scanning is unbypassable (industry-first)

Deep Reflexion (Inner-Outer Dual-Layer Reflective Architecture): Improves agent task success rate (PASS@1).

• Inner-to-Outer Failure Experience Conversion: Based on LLM-driven automatic root cause analysis (RCA) and heuristic strategy extraction, Reflexion is dynamically injected into the context space, achieving an inner-outer dual-layer collaborative closed-loop reflective correction and intelligent retry
• Inner Failure Prevention Mechanism: Three-layer failure prevention system (trajectory repetition, error action repetition, loop pseudo-termination) monitors failure risks and injects context, improving task success rate (PASS@1)

What Can You Do With OmniAgent

Quick Start

Requirements: Python 3.11+, an LLM API key (DeepSeek / OpenAI / Anthropic / Ollama / Gemini).

Installation

# 1. Install
$ pip install -e .

# 2. Interactive setup — choose provider, enter API key, done
$ omniagent onboard

# 3. Start
$ omniagent chat                    # CLI
$ omniagent serve                   # Web UI → http://127.0.0.1:18790

Three Ways to Interact

Configuration

Configuration is layered: defaults → ~/.omniagent/config.yaml → environment variables.

providers:
  deepseek:
    api_key: "sk-your-key"
    model_id: deepseek-chat

agent:
  model_provider: deepseek
  reflexion_enabled: true

Full configuration reference: docs.omniagent.dev (coming soon)

Architecture & Project Structure

…

omniagent/
├── agents/          # Core: reflexion loop, sentinel, guardian, skill/memory evolution, context management
├── security/        # Policy engine, approval, audit, sandbox
├── tools/           # Built-in tools
├── channels/        # Feishu, Discord, Telegram, Webhook
├── config/          # OmniAgentConfig + sub-configs
├── gateway/         # WebSocket + HTTP server
└── rl/              # GRPO + PRM training pipeline

Roadmap

Near-term

• Four-Layer Proactive Memory System (Proactive Memory 2.0) — Automatically extract and persist long-term memories from conversations, refine the proactive memory system, and resolve conflicts between self-evolution rules
• Agent Plan-Mode — Implement a new Agent planning mode that generates and confirms a plan before executing complex tasks, then executes step by step
• Channel Ecosystem Expansion — Add WeChat, WeCom, DingTalk, and more connectors; improve channel abstraction layer to reduce integration cost
• Multi-Agent Collaboration — Enhance inter-agent communication and task orchestration, support dynamic delegation and result aggregation between agents
• Documentation
• More Tests

Contributing

1. Fork the repo and create a feature branch
2. Make your changes with tests
3. Run python -m pytest to verify
4. Submit a pull request

We welcome contributions of all kinds — bug fixes, new tools, channel connectors, and documentation improvements.

License

This project is licensed under GPL-3.0. Any code that references this project must also be open-sourced under the same license.