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

ToolSelf introduces a runtime self-reconfiguration paradigm for LLM-powered agents that dynamically adapts task execution strategies during operation rather than relying on static pre-execution configurations. The approach unifies configuration updates with task execution through a standardized tool interface, achieving 28.8-point performance gains over static baselines after Configuration-Aware Two-stage Training.

ToolSelf addresses a fundamental limitation in current LLM-based agentic systems: the inability to adapt configurations during task execution. Traditional approaches force developers to choose between specialized high-performance agents with narrow scopes or generalist agents with broad capabilities but weaker performance—a trade-off that hampers real-world deployment where tasks are unpredictable and complex.

The research emerges from growing recognition that static agent architectures waste computational resources and miss critical feedback signals. Prior attempts to solve this through pre-execution optimization, hierarchical planning, or post-hoc patching remain disconnected from actual task execution, creating information loss and unclear responsibility for failures. ToolSelf's innovation lies in treating configuration changes as first-class actions within the agent's decision space, enabling seamless adaptation based on real-time task progress.

The Configuration-Aware Two-stage Training methodology combines rejection sampling fine-tuning with trajectory-level KTO reinforcement learning to teach agents when and how to self-reconfigure effectively. The 28.8-point average improvement demonstrates substantial performance gains, suggesting the paradigm resolves the generalization-specialization tension that has constrained agent capabilities.

For the AI industry, this work signals movement toward autonomous systems capable of meta-reasoning about their own configurations—a prerequisite for deploying agents in heterogeneous real-world environments. The framework's standardized tool interface provides a generalizable abstraction that other researchers can build upon, potentially accelerating progress in adaptive AI systems across domains requiring both flexibility and performance.

• →ToolSelf enables LLM agents to dynamically reconfigure sub-goals, strategies, and tool selections during execution rather than before task initiation.
• →The approach achieves 28.8-point average performance improvement over static-configuration agents through integrated execution and adaptation.
• →Configuration-Aware Two-stage Training combines rejection sampling and reinforcement learning to teach agents effective self-reconfiguration patterns.
• →Zero-shot ToolSelf already rivals task-specialized agents, suggesting the paradigm addresses core generalization-specialization trade-offs in agentic AI.
• →The research establishes a path toward emergent agent adaptivity without manual guidance injection, reducing engineering overhead for multi-domain deployment.