ToolSelf: Unifying Task Execution and Self-Reconfiguration via Tool-Driven Emergent Adaptation

The article introduces ToolSelf, a novel paradigm for LLM-powered agentic systems that integrates task execution and self-reconfiguration through a tool-driven approach. ToolSelf utilizes a standardized tool interface to enable dynamic updates of sub-goals and strategies based on real-time feedback, addressing the limitations of static configurations. The Configuration-Aware Two-stage Training (CAT) method enhances adaptability, achieving an average improvement of 28.8 points over static configurations across diverse benchmarks, which is significant for practitioners seeking to develop more flexible and efficient AI agents.