An end-to-end, self-evolving training framework that unifies self-questioning, self-navigating, and self-attributing into a cohesive system, empowering agents to autonomously improve their capabilities for efficient, cost-effective, and continuous capability evolution.
One-Minute Overview#
AgentEvolver is a revolutionary AI agent training framework that enables agents to learn, think, and improve autonomously just like humans. Through three core mechanisms (self-questioning, self-navigating, and self-attributing), AgentEvolver allows AI systems to continuously evolve in complex environments without human intervention.
Core Value: Significantly reduces AI training costs, improves learning efficiency, and enables agents to autonomously adapt and evolve.
Quick Start#
Installation Difficulty: Medium - Requires conda, cuda toolkit, and multi-step environment setup, but provides detailed installation guides
# Basic dependency installation
conda activate agentevolver
# Option 1: Minimal example without ReMe
python launcher.py --conf examples/basic.yaml --with-appworld
# Option 2: Full example with all self-evolving mechanisms
python launcher.py --conf examples/overall.yaml --with-appworld --with-reme
Is this suitable for my scenario?
- ✅ Research institutions/AI labs: Developing autonomous evolving agent systems
- ✅ Game AI development: Especially for multi-agent social reasoning games (like Avalon and Diplomacy)
- ✅ RL algorithm researchers: Developers needing efficient reinforcement learning frameworks
- ❌ Rapid prototyping: Simple AI applications requiring immediate deployment
- ❌ Limited computing resources: High requirements for environment configuration and hardware
Core Capabilities#
1. Automatic Task Generation (Self-Questioning)#
- Agents autonomously explore environments and create diverse tasks, eliminating costly manual dataset construction Actual Value: Saves over 90% of task design time, enabling systems to automatically discover valuable learning tasks
2. Experience-guided Exploration (Self-Navigating)#
- Agents summarize and reuse cross-task experience, guiding higher-quality rollouts and improving exploration efficiency Actual Value: Learning efficiency improved by over 50%, reducing trial-and-error iterations
3. Attribution-based Credit Assignment (Self-Attributing)#
- Agents process long trajectories to uncover causal contribution of intermediate steps, enabling fine-grained and efficient policy optimization Actual Value: More precise optimization processes, 30% faster training convergence
Tech Stack & Integration#
Development Language: Python Key Dependencies: ReMe (experience management system), veRL (distributed RL training), mkdocs (documentation) Integration Method: Framework/Library
Ecosystem & Extensions#
- Game Arena: Extended to multi-agent social game environments, providing web-based interaction, scalable evaluation, and end-to-end training support
- Modular Architecture: Decoupled components allow easy customization and secondary development, supporting future algorithm upgrades
- Environment Compatibility: Standardized interfaces for seamless integration with various external environments and tool APIs
Maintenance Status#
- Development Activity: Very active - From the news section, the project has major updates and releases monthly
- Recent Updates: New version and features released in December 2025
- Community Response: Regular updates and technical reports available, but community response level is unknown
Commercial & Licensing#
License: Apache-2.0
- ✅ Commercial: Commercial use allowed
- ✅ Modification: Modification and redistribution allowed
- ⚠️ Restrictions: Must preserve original license and copyright notices
Documentation & Learning Resources#
- Documentation Quality: Comprehensive
- Official Documentation: https://github.com/modelscope/AgentEvolver
- Example Code: Multiple complete examples available (basic.yaml, overall.yaml, etc.)