Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{208215,
        author = {L. Kishore and Shenbagavadivu S and Mathiyarasi S and Preethi R and Vaishnavi Devi J and Veeraselvi G},
        title = {Small Language Models for Autonomous AI Agents: A Systematic Review of Efficient Reasoning, Tool Use, Planning and Reliability},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {4},
        pages = {1074-1089},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=208215},
        abstract = {Autonomous AI agents built on large language models (LLMs) have demonstrated strong reasoning, planning and tool-use capabilities, but their computational cost, latency, energy consumption and centralized deployment increasingly conflict with the operational requirements of real-world agentic systems. Small language models (SLMs) have re-emerged as a candidate substrate for such systems, offering order-of-magnitude reductions in inference cost while retaining much of the narrow, task-scoped competence that agentic sub-tasks actually require. This paper presents a systematic review of the literature on SLMs for autonomous agents, critically synthesizing work across four capability axes—reasoning, planning, tool use and memory-augmented retrieval—alongside the efficiency, latency, energy and privacy considerations that motivate their adoption. Existing surveys treat SLM capability, agent architecture, model routing and verification as largely separate literatures; this review's principal contribution is a Reliability-Focused Taxonomy that unifies them by explicitly linking SLM Capability to Agent Role, to a characteristic Failure Mode, to an applicable Verification mechanism, and to an Adaptive Escalation decision. Under the resulting framework, easy, low-risk sub-tasks are executed directly by an SLM; uncertain sub-tasks are executed with an accompanying verification pass; and complex or high-risk sub-tasks are escalated to a larger model, with the escalation decision and its outcome logged for continual recalibration. We compare this framework against existing cascade, routing and self-verification approaches, identify concrete research gaps in cross-task transferability of verifiers, standardized escalation benchmarks and lifecycle energy accounting, and outline directions for future work.},
        keywords = {adaptive escalation, agentic AI, edge inference, model cascading, reliability, small language models, tool use, verification.},
        month = {September},
        }