Agentic AI for Autonomous Problem Solving

Abstract:

Large language models are increasingly deployed as autonomous agents that must plan, invoke external tools, and revise their own approach until a real-world problem is solved, rather than simply answering a single-turn question. Reasoning-only techniques such as chain-of-thought prompting and tree-structured search improve internal deliberation but do not address the distinct requirement of acting on an external environment, while unconditional self-correction mechanisms have been shown in prior work to sometimes degrade an already-correct answer when no external verification signal is available. This paper proposes a hybrid agentic AI architecture for autonomous problem solving that combines an LLM-based planner and task decomposer, a tool-use and action execution layer, episodic and semantic memory, and a self-verification and reflection loop that triggers corrective replanning only when an explicit, externally groundable check against the goal state fails. Drawing on established agentic-reasoning, tool-use, self-correction, and agent-evaluation literature, the study defines seven evaluation criteria — Task Success Rate, Plan-Execution Precision, Subgoal Recall, F1 Score, Hallucinated-Action Rate, verification-layer AUC-ROC, and a combined Agentic Problem-Solving Reliability Score — and applies them to an illustrative case study of autonomous, multi-step problem solving. A five-configuration ablation study, spanning a naive single-shot baseline through a deliberately over-engineered configuration that adds unconditional multi-agent debate, is used to identify the point at which additional correction machinery stops helping and begins to erode reliability, and the recommended configuration is benchmarked against the naive baseline under a stress test of complex, multi-step tasks requiring tool composition. The illustrative results indicate that the proposed configuration achieves both the highest task-solving F1 score and the lowest hallucinated-action rate among all evaluated configurations, with the largest relative improvement over the naive baseline occurring on the most complex, multi-step tasks. The findings position externally grounded, selectively triggered self-verification, rather than the unconditional stacking of every available reasoning or correction mechanism, as a practical design principle for deploying reliable autonomous problem-solving agents.