Paper deep dive
Don't Overthink, Don't Underthink: Toward Adaptive Reasoning in Agentic AI
Md Jueal Mia, M. Hadi Amini
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 8/28/2026, 3:44:46 AM
Summary
This paper investigates reasoning efficiency in agentic AI systems, identifying over-reasoning and under-reasoning as critical failure modes. Through empirical evaluation on MATH-500 and GAIA benchmarks using models like Phi-4-reasoning, Qwen3.5-4B, and Llama-3.1-8B-Instruct, the authors demonstrate that excessive reasoning leads to high computational costs without proportional accuracy gains, while insufficient reasoning results in incorrect or incomplete solutions. The study advocates for adaptive reasoning mechanisms that dynamically allocate computational resources based on evolving task demands.
Entities (12)
Relation Signals (8)
M. Hadi Amini → affiliatedwith → Florida International University
confidence 99% · M. Hadi Amini... Florida International University
Md Jueal Mia → affiliatedwith → Florida International University
confidence 99% · Md Jueal Mia... Florida International University
Phi-4-Reasoning → evaluatedon → MATH-500
confidence 98% · Each model was evaluated on all 500 MATH-500 test problems
Phi-4-Reasoning → evaluatedon → GAIA
confidence 98% · Each model was evaluated on... 165 GAIA validation tasks
Over-reasoning → causes → High Computational Cost
confidence 96% · cases classified as over-reasoning are associated with higher computational cost without proportional accuracy gains
Under-reasoning → causes → Incorrect solutions
confidence 96% · cases classified as under-reasoning are consistently associated with incorrect or incomplete solutions
Phi-4-Reasoning → exhibitshighoverreasoningrate → Over-reasoning
confidence 95% · Phi-4-reasoning exhibits the highest over-reasoning rate (89.40%)
LangGraph → usedin → Agentic AI
confidence 95% · quantitative analysis of reasoning allocation in a LangGraph-based agentic AI framework
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent advances in Large Language Models (LLMs) have shown that increased inference-time reasoning can improve performance on complex tasks. However, many existing approaches rely on fixed or preallocated reasoning controls, such as fixed token budgets, pre-execution difficulty estimates, or activation-space interventions, and are often evaluated on standalone reasoning benchmarks rather than full agentic workflows. These assumptions may not hold in agentic AI systems, where reasoning requirements evolve dynamically through planning, tool use, memory retrieval, and agent-to-agent interactions. Consequently, reasoning can become either excessive or insufficient, resulting in unnecessary computation, increased latency, planning drift, excessive tool use, or incomplete solutions. We argue that a major challenge for next-generation agentic AI is not merely how much reasoning a language model should perform, but how it should allocate reasoning according to evolving task demands. We characterize over-reasoning and under-reasoning as recurring failure modes of misallocated reasoning and evaluate them on MATH-500 and the GAIA public validation benchmark. Using tool-decision latency, token consumption, token-limit exhaustion, and answer correctness, our results suggest that cases classified as over-reasoning are associated with higher computational cost without proportional accuracy gains, whereas cases classified as under-reasoning are consistently associated with incorrect or incomplete solutions. These findings motivate future research on adaptive reasoning mechanisms for agentic AI.
Tags
Links
- Source: https://arxiv.org/abs/2608.26442v1
- Canonical: https://arxiv.org/abs/2608.26442v1
Trouble viewing inline? Open PDF directly →
Full Text
30,553 characters extracted from source content.
Expand or collapse full text
Don’t Overthink, Don’t Underthink: Toward Adaptive Reasoning in Agentic AI Md Jueal Mia and M. Hadi Amini mmia001, moamini@fiu.edu Knight Foundation School of Computing and Information Sciences, Security, Optimization, and Learning for InterDependent Networks Laboratory (solid lab) Florida International University Miami, Florida, USA Abstract Recent advances in Large Language Models (LLMs) have shown that increased inference-time reasoning can improve performance on complex tasks. However, many existing approaches rely on fixed or preallocated reasoning controls, such as fixed token budgets, pre- execution difficulty estimates, or activation-space interventions, and are often evaluated on standalone reasoning benchmarks rather than full agentic workflows. These assumptions may not hold in agentic AI systems, where reasoning requirements evolve dynam- ically through planning, tool use, memory retrieval, and agent- to-agent interactions. Consequently, reasoning can become either excessive or insufficient, resulting in unnecessary computation, increased latency, planning drift, excessive tool use, or incomplete solutions. We argue that a major challenge for next-generation agentic AI is not merely how much reasoning a language model should perform, but how it should allocate reasoning according to evolving task demands. We characterize over-reasoning and under- reasoning as recurring failure modes of misallocated reasoning and evaluate them on MATH-500 and the GAIA public validation benchmark. Using tool-decision latency, token consumption, token- limit exhaustion, and answer correctness, our results suggest that cases classified as over-reasoning are associated with higher com- putational cost without proportional accuracy gains, whereas cases classified as under-reasoning are consistently associated with in- correct or incomplete solutions. These findings motivate future research on adaptive reasoning mechanisms for agentic AI. Keywords agentic AI, adaptive reasoning, over-reasoning, under-reasoning, reasoning efficiency 1 Introduction Recent advancements in LLMs have significantly enhanced their reasoning capabilities, enabling them to solve complex problems through step-by-step deliberation. This progress has led to the emergence of a class of models known as Large Reasoning Models (LRMs), which are typically developed through supervised fine- tuning (SFT) and reinforcement learning from human feedback (RLHF). By generating chain-of-thought (CoT) reasoning traces, LRMs can perform sophisticated tasks involving logical deduction, mathematical reasoning, planning, and decision-making [12]. As a result, these models have demonstrated strong performance across a wide range of challenging benchmarks. These advances build upon a series of foundational developments, including Chain-of-Thought prompting for explicit reasoning [16], reasoning-and-acting frame- works such as ReAct [18] and the emergence of dedicated reasoning models including OpenAI o1 [9] and Phi-4-reasoning [1]. Despite these improvements, explicit reasoning introduces sub- stantial computational overhead during inference by increasing latency and token consumption, leading to higher inference costs. Recent surveys identify efficient inference as a key challenge for next-generation language models, emphasizing the need for mecha- nisms that better allocate computational resources during inference [4]. Although extensive reasoning is often beneficial for complex tasks, it is frequently unnecessary for simpler queries, and exist- ing LRMs often allocate reasoning inefficiently by over-computing on easy problems while under-reasoning on more difficult ones, creating a trade-off between reasoning quality and computational efficiency [3]. This challenge becomes even more pronounced in agentic AI systems, where one or more LLMs coordinate planning, tool use, memory retrieval, environment interaction, and multi- step decision making [10]. In such systems, specialized components (e.g., planners and response generators) may independently invoke LLMs and perform their own reasoning processes, causing unnec- essary reasoning at individual components to accumulate across the pipeline and significantly increase overall latency, token usage, and computational cost. As inference-time computation becomes a important factor in the deployment of modern AI systems, understanding how to dy- namically adapt reasoning effort to the complexity of a given task has emerged as an important research challenge. Developing mech- anisms that allocate reasoning resources more efficiently can reduce unnecessary computational costs while maintaining, or even im- proving, overall task performance. This challenge is particularly im- portant for agentic AI systems, where reasoning efficiency directly impacts scalability, responsiveness, and real-world usability. Fig- ure 1 illustrates the major challenges that arise from mis-allocated reasoning that motivate this work. The main contributions of our work are listed below. •We investigate the impact of over-reasoning and under- reasoning in agentic AI, showing how reasoning effort that is either excessive or insufficient relative to task require- ments is associated with higher inference cost or degraded task performance. •We present a quantitative analysis of reasoning allocation in a LangGraph-based agentic AI framework using repre- sentative reasoning models on MATH-500 and GAIA. Our results reveal that current reasoning models frequently 1 arXiv:2608.26442v1 [cs.AI] 26 Aug 2026 Mia and Amini LLM Agent Tools Data Search Web Search Memory Prompt LLM Agent User Query +Tools Observations Response Redundant Reasoning Repeated Tool Usage Figure 1: Overview of over-reasoning and under-reasoning in agentic AI. The agent loop is annotated with four failure modes: redundant reasoning, weak reasoning, repeated tool usage, and no tool usage (skipped tool calls). exhibit over-reasoning or under-reasoning, resulting in in- efficient reasoning allocation for agentic AI. •Based on our findings, we identify key research challenges and outline candidate evaluation dimensions and future directions for adaptive reasoning mechanisms that dynami- cally determine when to reason, when to act, and when to stop. 2 Related Works Recent studies have explored the growing role of reasoning in agen- tic AI systems. Wu et al. [17] demonstrated that integrating search, coding, and memory agents can enhance complex reasoning capa- bilities . As agentic reasoning becomes increasingly sophisticated, researchers have begun examining its efficiency implications. Kim et al. [7] showed that deeper reasoning and test-time scaling intro- duce substantial computational and infrastructure costs. Similarly, Tran et al. [14] found that increasing reasoning and coordination does not always improve performance, with single-agent systems of- ten outperforming multi-agent architectures under equal compute budgets. Wang et al. [15] showed that increasing reasoning budgets, adding more planning steps, and employing complex memory mech- anisms often result in only marginal performance improvements while significantly increasing computational cost, while Becker et al. [5] revealed that prolonged agent interactions can cause discus- sions to drift away from task objectives, leading to performance degradation. Collectively, these findings suggest that increasing reasoning effort, agent interactions, and test-time computation does not necessarily translate into better agent performance and may instead introduce inefficiencies such as overthinking, planning drift, and coordination overhead. Despite this progress, many existing approaches still rely on fixed reasoning budgets or evaluate adaptive reasoning outside full agentic workflows. These assumptions may not hold in agentic AI, where reasoning requirements evolve dynamically through in- teractions with tools, memory, and external environments. As a result, agents must determine when to continue reasoning, invoke tools, or act, while additional deliberation may amplify noisy obser- vations rather than improve decisions. Moreover, prior work has shown that prolonged reasoning and replanning can lead to plan- ning drift, where agents gradually deviate from the user’s objective, although planning drift is not directly measured in this study. These observations raise a key question: when does additional reasoning improve agent performance, and when does it become counterpro- ductive? Answering this question is essential for developing agentic AI systems that allocate reasoning efficiently while maintaining alignment with task objectives. 3 Preliminaries 3.1 Agentic AI Systems Agentic AI represents a shift from traditional AI systems toward au- tonomous, goal-oriented systems capable of operating in dynamic environments with minimal human intervention [2]. Unlike conven- tional AI, which is typically designed for predefined tasks, agentic AI can adapt to changing conditions, make decisions based on con- textual information, and pursue complex objectives over extended periods [6]. Recent surveys further characterize agentic AI as an evo- lution beyond single-agent architectures, emphasizing multi-agent collaboration, dynamic task decomposition, persistent memory, and coordinated autonomy to accomplish complex goals across diverse application domains [10]. These capabilities enable agentic systems to move beyond simple automation toward autonomous problem solving, adaptive decision-making, and collaborative task execution in real-world environments. 3.2Under-Reasoning, Adequate Reasoning, and Over-Reasoning Recent advances in reasoning models show that increasing reason- ing effort does not always improve performance. Models may allo- cate reasoning inefficiently, producing unnecessarily long reasoning traces for simple problems while failing to devote sufficient reason- ing to more challenging ones, and answer correctness often exhibits a non-linear relationship with reasoning length [11]. In agentic set- tings, stronger reasoning improves planning-oriented tasks but incurs substantially higher computational costs and reasoning over- head [20]. Moreover, excessive test-time reasoning yields dimin- ishing returns and may even cause models to abandon previously correct answers, highlighting the risks of overthinking during in- ference [19]. These findings suggest that agentic systems should 2 Don’t Overthink, Don’t Underthink: Toward Adaptive Reasoning in Agentic AI adapt reasoning effort to task complexity rather than relying on uniformly long reasoning processes. Conceptually, we distinguish three reasoning regimes. Under- reasoning occurs when an agent allocates insufficient reasoning, evidence gathering, or verification for the task, such as omitting intermediate reasoning steps, failing to gather sufficient evidence, or terminating prematurely. Over-reasoning occurs when an agent continues reasoning beyond what is necessary, exhibiting redun- dant deliberation, repeated verification, or unnecessary reasoning that increases computational cost. Between these extremes lies ad- equate reasoning, where reasoning effort is sufficient for the task without exhibiting either under-reasoning or over-reasoning. 4 Preliminary Study: Evidence of Misallocated Reasoning 4.1 Experimental Setup Experiments were conducted on a server with two NVIDIA RTX A6000 GPUs (48,GB VRAM each). Our LangGraph-based agentic AI framework was served through vLLM-compatible APIs using Qwen3.5-4B as the router, 1 while the final-response models were Qwen3.5-4B, Llama-3.1-8B-Instruct, 2 and Phi-4-reasoning. 3 The Qwen3.5-4B configuration used the same model for both router and final response. Each model was evaluated on all 500 MATH-500 test problems [13] and 165 GAIA validation tasks [8] using deterministic decoding (temperature=0.0) with a maximum generation length of 4,096 tokens. We recorded tool usage, latency, token consumption, token-limit hits, and complete interaction traces. For reasoning models, reasoning tokens were computed from tokens enclosed by<think>and</think>; since Llama-3.1-8B-Instruct does not expose reasoning traces, this metric is reported asN/A. Finally, GPT-4.1 independently classified each trajectory as over-reasoning, under-reasoning, or adequate reasoning, while final answers were labeled as correct, incorrect, or incomplete, with reasoning labels assigned independently of answer correctness. 5 Findings Analysis of Tables 1 and 2 suggests that both over-reasoning and under-reasoning emerge as recurring failure modes in the evaluated LangGraph-based agent across reasoning-intensive (MATH-500) and real-world multi-step (GAIA) tasks. Rather than consistently allocating reasoning effort according to task requirements, the eval- uated agent frequently allocates either excessive or insufficient reasoning, resulting in increased computational cost or degraded task performance. The first major issue is over-reasoning. On MATH-500, Phi- 4-reasoning exhibits the highest over-reasoning rate (89.40%), fol- lowed by Qwen3.5-4B (68.00%) and Llama-3.1-8B-Instruct (19.80%). It also generates the most reasoning tokens (2220.12), the longest outputs (4046.59 tokens), and the highest final inference time (180.80 s), compared with Qwen3.5-4B (1732.39 reasoning tokens, 2461.67 output tokens, 70.21 s) and Llama-3.1-8B-Instruct (904.48 output tokens, 21.42 s). Although Phi-4-reasoning achieves the highest accuracy (92.40%), it reaches the maximum token limit in 482 of 1 Hugging Face identifier: Qwen/Qwen3.5-4B. 2 Hugging Face identifier: meta-llama/Llama-3.1-8B-Instruct. 3 Hugging Face identifier: microsoft/Phi-4-reasoning. 500 samples, compared with 190 and 71 for Qwen3.5-4B and Llama- 3.1-8B-Instruct, respectively. A similar trend is observed on GAIA, where Phi-4-reasoning again produces the longest reasoning traces (2897.95 reasoning tokens, 4091.12 output tokens, 182.94 s). Since Llama-3.1-8B-Instruct does not expose explicit reasoning traces, its reasoning-token count is reported asN/A. Overall, increasing explicit reasoning substantially increases inference cost. The second major issue is that over-reasoning does not nec- essarily translate into proportional performance gains. On MATH-500, Phi-4-reasoning improves accuracy over Qwen3.5-4B (92.40% vs. 87.60%) but nearly triples final inference time (180.80 s vs. 70.21 s), increases output length (4046.59 vs. 2461.67 tokens), and more than doubles token-limit hits (482 vs. 190). On GAIA, despite producing much longer reasoning traces and requiring almost three times longer inference than Qwen3.5-4B (182.94 s vs. 67.49 s), Phi- 4-reasoning achieves slightly lower accuracy (11.52% vs. 12.12%). These observations indicate that substantially increasing reasoning effort does not necessarily yield proportional performance improve- ments. The third major issue is under-reasoning. On MATH-500, Llama- 3.1-8B-Instruct exhibits the highest under-reasoning rate (22.00%), followed by Qwen3.5-4B (8.00%) and Phi-4-reasoning (4.80%), result- ing in 64, 48, and 33 incomplete responses, respectively. On GAIA, under-reasoning increases to 71.52% for Llama-3.1-8B-Instruct, 62.42% for Phi-4-reasoning, and 36.97% for Qwen3.5-4B, corresponding to 56, 130, and 66 incomplete responses. Although reasoning cate- gories are assigned independently of correctness, all under-reasoning cases were associated with incorrect or incomplete answers, sug- gesting that reasoning often terminates before collecting sufficient evidence or adequately verifying conclusions. The fourth major issue is token budget exhaustion. On MATH- 500, Phi-4-reasoning reaches the maximum output token limit in 482 of 500 samples (96.40%), substantially exceeding Qwen3.5-4B (190; 38.00%) and Llama-3.1-8B-Instruct (71; 14.20%). Despite this, its hit accuracy remains high (92.12%), indicating that many responses are already correct before reaching the limit. On GAIA, Phi-4-reasoning again records the most token-limit hits (164 of 165), compared with 75 and 10 for Qwen3.5-4B and Llama-3.1-8B-Instruct, while hit accuracy remains low for both reasoning models (11.59% and 13.33%), suggesting that exhausting the token budget does not substantially improve performance on complex tasks. The fifth major issue is task-dependent reasoning and tool- use allocation. On MATH-500, the agent averages 0.39 tool calls per sample, consuming approximately 583 tool-input tokens and 1225 tool-output tokens despite external information rarely be- ing required. In contrast, GAIA averages 0.93 tool calls per sam- ple because external information and multi-step interactions are frequently necessary. Nevertheless, despite greater tool use, the highest GAIA accuracy remains only 12.12%, suggesting that in- creasing reasoning length and tool interactions alone is insufficient to substantially improve task performance. Overall, the evaluated LangGraph-based agent does not consistently allocate reasoning effort or tool use efficiently across different task types. 3 Mia and Amini Table 1: Average runtime and reasoning statistics in agentic AI across MATH-500 and GAIA. Tool agent model: Qwen3.5-4B. DatasetFinal ModelAvg. Total Time (s) Avg. Tool Decision Time (s) Avg. Final Answer Time (s) Avg. Tool Calls/Sample Avg. Tool Input Tokens Avg. Tool Output Tokens Avg. Final Output Tokens Avg. Final Reasoning Tokens Token Limit Hits MATH-500 Qwen3.5-4B105.7634.9570.2100.39583.301224.892461.671732.39190 Llama-3.1-8B-Instruct57.7635.7521.4200.39583.391225.09904.48N/A71 Phi-4-reasoning216.4635.08180.8000.39583.301224.894046.592220.12482 GAIA Qwen3.5-4B79.6410.9267.4900.93582.24379.532353.901219.7675 Llama-3.1-8B-Instruct22.7710.9410.7100.93582.24379.53452.23N/A10 Phi-4-reasoning195.0010.74182.9400.93582.24379.534091.122897.95164 Table 2: Reasoning efficiency and answer quality in agentic AI across MATH-500 and GAIA. Tool agent model: Qwen3.5-4B. Judge model: GPT-4.1. Hit Accuracy denotes accuracy on samples where the final model reached the maximum output token limit, while No-Hit Accuracy denotes accuracy on samples that did not reach the maximum output token limit. DatasetFinal ModelAccuracy(%) Answer Correct Answer Incorrect Incomplete Token Limit Hits Hit Accuracy(%) No-Hit Accuracy(%) Over-Reasoning Under-Reasoning Adequate Reasoning Over-Reasoning Rate(%) Under-Reasoning Rate(%) MATH-500 Qwen3.5-4B87.60438144819073.6896.133404012068.008.00 Llama-3.1-8B-Instruct49.8024918764719.8656.419911029119.8022.00 Phi-4-reasoning92.4046253348292.12100.00447242989.404.80 GAIA Qwen3.5-4B12.122079667513.3311.1163614138.1836.97 Llama-3.1-8B-Instruct5.45910056100.005.81171183010.3071.52 Phi-4-reasoning11.52191613016411.590.0058103435.1562.42 6 Discussion and Future Directions Our findings show that reasoning efficiency remains a key challenge for both current LRMs and agentic AI systems. On simpler reason- ing tasks, reasoning models can improve accuracy but often incur substantially higher latency, token consumption, and token-budget exhaustion. On more complex agentic tasks, additional reasoning greatly increases computational cost without corresponding accu- racy gains, while insufficient reasoning often results in incorrect or incomplete solutions. These challenges become more pronounced in agentic AI, where reasoning is intertwined with planning, tool use, memory retrieval, and multi-step decision making. Together, these observations motivate four key research challenges:1Unknown Reasoning Sufficiency, where agents cannot determine when enough reasoning has been performed and should stop, leading to unnecessary computation;2Over-Reasoning and Planning Drift, where excessive reasoning increases computational cost and may introduce redundant deliberation, planning drift, or incor- rect intermediate reasoning without improving answer quality; 3 Under-Reasoning and Premature Decisions, where agents fail to recognize when additional reasoning, evidence gathering, or verification is required before producing a final response; and 4Dynamic Reasoning and Tool Allocation, where reasoning requirements evolve throughout execution as new observations are obtained from tools, memory, or the environment, requiring adap- tive mechanisms that jointly allocate reasoning effort, tool usage, and token budgets according to task complexity while avoiding both over- and under-reasoning. Future research should focus on adaptive reasoning mechanisms that dynamically determine when to reason, act, and stop by balanc- ing reasoning cost, uncertainty, task complexity, and the expected benefit of additional computation. A Candidate Adaptive-Reasoning Controller. One promising di- rection is the development of an adaptive-reasoning controller that dynamically allocates reasoning effort throughout an agent’s execu- tion. Rather than relying on fixed reasoning budgets, the controller would determine whether to continue reasoning, invoke external tools, or produce a final answer based on the current reasoning state, task requirements, and the expected benefit of additional computation. Such a controller could balance reasoning quality against computational cost, enabling agents to allocate reasoning only when it is likely to improve task performance. A Measurable Evaluation Suite. More broadly, reasoning-efficient agentic AI will require new benchmarks, evaluation metrics, and control mechanisms that explicitly measure and optimize reasoning allocation rather than simply increasing reasoning effort. Progress toward adaptive reasoning therefore requires evaluation beyond task accuracy alone. Future work could consider an evaluation suite that measures how effectively reasoning effort is allocated, includ- ing reasoning efficiency (accuracy per reasoning token), stopping behavior, unnecessary tool usage, planning drift, and computational cost. Such metrics could enable systematic evaluation of reasoning allocation strategies and reveal trade-offs that are not captured by answer accuracy alone. 7 Conclusion In this paper, we investigated reasoning efficiency in agentic AI and identified over-reasoning and under-reasoning as recurring failure modes of reasoning allocation. Our findings suggest that cases classified as over-reasoning are associated with higher latency, greater token consumption, more frequent tool usage, and increased token-limit exhaustion without proportional improvements in task performance, whereas cases classified as under-reasoning are con- sistently associated with incorrect or incomplete solutions under our evaluation rubric. Although MATH-500 and GAIA exhibit dif- ferent dominant failure modes, both benchmarks indicate that the evaluated agent configuration does not always allocate reasoning effort and tool usage according to task requirements. These findings suggest that the goal of future agentic AI systems should not be to maximize reasoning effort, but to allocate reasoning adaptively according to evolving task demands. We hope this work motivates future research on adaptive reasoning mechanisms that dynami- cally determine when to reason, when to act, and when to stop. 8 Limitations This work presents a preliminary empirical study of reasoning allo- cation in a LangGraph-based agentic AI framework using a fixed Qwen3.5-4B tool-routing model and three final-response models. Consequently, our findings should not be generalized to all agentic architectures, routing strategies, or foundation models. Although reasoning regimes are independently classified from answer correct- ness, our analysis does not explicitly distinguish tool-routing fail- ures, retrieval failures, reasoning failures, response synthesis errors, 4 Don’t Overthink, Don’t Underthink: Toward Adaptive Reasoning in Agentic AI or formatting errors, all of which may contribute to incorrect predic- tions. We evaluate inference cost using latency, token consumption, tool usage, and token-limit hits, but do not measure monetary cost, GPU utilization, energy consumption, or throughput. Furthermore, our experiments use a fixed maximum generation length of 4,096 tokens and do not include token-budget ablations, no-tool baselines, or standalone LLM baselines. Finally, we report average statistics only; future work should incorporate confidence intervals, distribu- tional analyses, qualitative trace studies, and broader evaluations across additional agentic frameworks, models, and tasks. Ethics and Reproducibility Statement This work investigates the reasoning challenges that emerge when LLMs are integrated into agentic AI systems. Our analysis identifies over-reasoning and under-reasoning as recurring failure modes that can increase computational cost or degrade task performance, motivating future research on more adaptive reasoning mecha- nisms. To support reproducibility, we will release the complete source code for our LangGraph-based framework, experimental configurations, prompts (including the GPT-4.1 judge prompt and reasoning-classification rubric), tool definitions, and per-sample evaluation outputs. Acknowledgement This work is partly based upon the work supported by the National Center for Transportation Cybersecurity and Resiliency (TraCR) (a U.S. Department of Transportation National University Transporta- tion Center) headquartered at Clemson University, Clemson, South Carolina, USA. Any opinions, findings, conclusions, and recommen- dations expressed in this material are those of the author(s) and do not necessarily reflect the views of TraCR, and the U.S. Government assumes no liability for the contents or use thereof. This research is also supported by the National Artificial Intel- ligence Research Resource (NAIRR) Pilot and AWS through the CloudBank project, which is supported by National Science Foun- dation grant#1925001. Any opinions, findings, conclusions, and recommendations expressed in this material are those of the au- thor(s) and do not necessarily reflect the views of NAIRR, and the U.S. Government assumes no liability for the contents or use thereof. References [1] Marah Abdin, Sahaj Agarwal, Ahmed Awadallah, Vidhisha Balachandran, Harki- rat Behl, Lingjiao Chen, Gustavo de Rosa, Suriya Gunasekar, Mojan Java- heripi, Neel Joshi, et al.2025. Phi-4-reasoning technical report. arXiv preprint arXiv:2504.21318 (2025). [2]Deepak Bhaskar Acharya, Karthigeyan Kuppan, and B Divya. 2025. Agentic AI: Autonomous intelligence for complex goals—A comprehensive survey. IEEe Access 13 (2025), 18912–18936. [3]Pranjal Aggarwal, Seungone Kim, Jack Lanchantin, Sean Welleck, Jason Weston, Ilia Kulikov, and Swarnadeep Saha. 2025. Optimalthinkingbench: Evaluating over and underthinking in llms. arXiv preprint arXiv:2508.13141 (2025). [4]Hadi Amini, Md Jueal Mia, Yasaman Saadati, Ahmed Imteaj, Seyedsina Nabavi- razavi, Urmish Thakker, Md Zarif Hossain, Awal Ahmed Fime, and S Iyengar. 2025. Distributed llms and multimodal large language models: A survey on advances, challenges, and future directions. arXiv preprint arXiv:2503.16585 (2025). [5]Jonas Becker, Lars Benedikt Kaesberg, Andreas Stephan, Jan Philip Wahle, Terry Ruas, and Bela Gipp. 2026. Stay focused: Problem drift in multi-agent debate. In Findings of the Association for Computational Linguistics: EACL 2026. 5068–5102. [6]Soodeh Hosseini and Hossein Seilani. 2025. The role of agentic ai in shaping a smart future: A systematic review. Array 26 (2025), 100399. [7]Jiin Kim, Byeongjun Shin, Jinha Chung, and Minsoo Rhu. 2026. The cost of dynamic reasoning: Demystifying ai agents and test-time scaling from an ai infrastructure perspective. In 2026 IEEE International Symposium on High Perfor- mance Computer Architecture (HPCA). IEEE, 1–16. [8] Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2024. Gaia: a benchmark for general ai assistants. In International Conference on Learning Representations, Vol. 2024. 9025–9049. [9]OpenAI. 2024. Learning to Reason with LLMs. Technical Report. OpenAI. https: //openai.com/index/learning-to-reason-with-llms/ [10]Ranjan Sapkota, Konstantinos I Roumeliotis, and Manoj Karkee. 2025. Ai agents vs. agentic ai: A conceptual taxonomy, applications and challenges. Information Fusion (2025), 103599. [11]Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. 2025. Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms. arXiv preprint arXiv:2505.00127 (2025). [12] Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, et al.2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419 (2025). [13] Hugging Face H4 Team. 2024. MATH-500. https://huggingface.co/datasets/ HuggingFaceH4/MATH-500. 500-problem evaluation subset of the MATH bench- mark. [14]Dat Tran and Douwe Kiela. 2026. Single-agent llms outperform multi-agent systems on multi-hop reasoning under equal thinking token budgets. arXiv preprint arXiv:2604.02460 (2026). [15] Ningning Wang, Xavier Hu, Pai Liu, He Zhu, Yue Hou, Heyuan Huang, Shengyu Zhang, Jian Yang, Jiaheng Liu, Ge Zhang, et al.2025. Efficient agents: Building effective agents while reducing cost. arXiv preprint arXiv:2508.02694 (2025). [16]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al.2022. Chain-of-thought prompting elicits reason- ing in large language models. Advances in neural information processing systems 35 (2022), 24824–24837. [17] Junde Wu, Jiayuan Zhu, Yuyuan Liu, Min Xu, and Yueming Jin. 2025. Agentic rea- soning: A streamlined framework for enhancing llm reasoning with agentic tools. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 28489–28503. [18] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629 (2022). [19]Shu Zhou, Rui Ling, Junan Chen, Xin Wang, Tao Fan, and Hao Wang. 2026. When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling. arXiv preprint arXiv:2604.10739 (2026). [20]Xueyang Zhou, Guiyao Tie, Guowen Zhang, Weidong Wang, Zhigang Zuo, Di Wu, Duanfeng Chu, Pan Zhou, Neil Zhenqiang Gong, and Lichao Sun. 2025. Exploring the Necessity of Reasoning in LLM-based Agent Scenarios. arXiv preprint arXiv:2503.11074 (2025). 5