An agent framework that evolves its own skills and brain online

What it does

OmniAgent is a Python agent framework that wraps LLM inference in a loop with tools, memory, and multi-channel chat support. It attempts to improve itself during conversation by auto-generating and repairing skills, updating user context from feedback, and tuning its BrainModel via an online reinforcement-learning loop. A four-layer security scan and dedicated planning and safety sub-agents are meant to keep the loop from going off the rails.

The interesting bit

The unusual pitch is full-dimensional “self-evolution”: not just prompt engineering, but native skill generation from high-frequency action patterns, progressive context loading staged by conversation depth, and GRPO + PRM reinforcement learning applied to the model itself during interaction. It is essentially trying to turn a static LLM wrapper into a continuously patching autonomous system.

Key highlights

• Claims real-time skill evolution: auto-creates, inspects, and repairs skills during execution rather than after.
• Proactive memory via a multi-layer stack that mixes explicit user feedback with implicit LLM induction to update preferences.
• Hyper-Harness scaffold with dynamic multi-agent dispatch (Sentinel for planning, Guardian for safety) and concurrent tool execution with dependency resolution.
• Four-layer security flow: LLM review, policy engine, interactive approval, and execution sandbox, with trust-level policies per skill.
• Supports multiple backends (OpenAI, Claude, DeepSeek, Ollama, Gemini) and channels (CLI, Web, Feishu, Discord, Telegram).

Caveats

• Documentation is still marked “on the way” and “coming soon,” so many details remain unspecified.
• The README makes strong comparative claims (e.g., “unbypassable” security, 90% token savings) without providing reproducible benchmarks or external evaluations.

Verdict

Worth a look if you are experimenting with agent architectures that blur the line between runtime scaffolding and online model adaptation. Skip it if you need mature docs, proven safety guarantees, or a simple, static ReAct wrapper.