APEX: A Three-Layer Co-Evolution Framework Enables True Self-Evolution of AI Agents

The APEX (Adaptive Principle EXtraction) framework addresses the limitations of existing single-dimensional self-improvement by simultaneously optimizing three dimensions: prompt templates, behavioral principles, and workflow topology. This framework achieves a 90% performance improvement in the production environment of the NVIDIA Agent Challenge, demonstrating the superiority of multi-dimensional co-evolution.

The current advanced Self-Harness framework only optimizes the single dimension of prompt templates. Although it achieves a 14-21% improvement in Terminal-Bench-2.0, its overall performance is limited because it does not involve the optimization of behavioral principles and workflow topology. It's like a team updating the operation manual but not changing the employees' thinking habits and collaboration processes—ultimately, the effect is greatly reduced.

The core of the APEX framework is the simultaneous evolution of three interrelated dimensions:

1. Prompt Template Optimization (L1)：Analyze failure mode clusters and target weak points in the template for repair;
2. Behavioral Principle Evolution (L₂)：Extract 6 novel and reusable principles from past successful execution records using successful trajectory distillation technology;
3. Workflow Topology Optimization (L3)：Automatically select the optimal workflow topology based on the structural fitness selection mechanism (e.g., the score of research-priority topology increased by 20%).

APEX has been validated in a production environment: deployed on Joe, an agent based on NVIDIA Nemotron (managing a 15-node cluster), it evolved using 114 real task trajectories over 18 days. The results show that the APEX health score increased from 0.3 to 0.57 (a 90% improvement), and it only required about 4 LLM calls (local qwen2.5-coder:32b) taking 270 seconds—low cost and efficient.

• Successful Trajectory Distillation: Identify key decision patterns from the agent's successful task trajectories and abstract them into general behavioral principles (e.g., code review prioritizes verifying interface contract consistency);
• Structural Fitness Selection: Adopt a genetic algorithm-style strategy to evaluate the efficiency and success rate of candidate workflow topologies and automatically retain the optimal structure (e.g., research-type tasks prioritize extensive search before in-depth analysis).

The APEX framework verifies the superiority of multi-dimensional co-evolution. Evolution based on real data is more reliable, and its efficient mechanism supports continuous self-improvement in production environments. Future research directions include: incorporating more dimensions (tool selection, memory management), cross-task principle migration, and developing more efficient evolution algorithms to promote continuous self-learning of AI agents.