PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolution of AI in Education: Agentic Workflows

Firuz Kamalov, David Santandreu Calonge, Linda Smail, Dilshod Azizov, Dimple R. Thadani, Theresa Kwong, Amara Atif

arXiv:2504.20082v1cs.AIcs.CY

TL;DR

Conventional LLMs face limits in adaptability, reasoning, and reliance on static training data. This paper reviews four agentic paradigms in education and presents a multi-agent essay-scoring proof of concept; its MASS system reports a lower MAE than stand-alone LLMs. The authors identify further work on trust, fairness, and educational implementation challenges.

  • Problem

    Conventional LLMs are constrained by static training data, limited adaptability, and limited nuanced reasoning, while educational agents require careful assessment across learning contexts.

    Source: §1 · p. 2

  • Method

    The paper reviews reflection, planning, tool use, and multi-agent collaboration and presents a proof-of-concept multi-agent essay-scoring system.

    Source: §1 · p. 2, §6 · p. 38

  • Results

    MASS achieved a mean absolute error (MAE) of 0.5612, compared with 0.6129 for stand-alone GPT-4o, 0.7345 for DeepSeek 67b, and 1.6956 for DeepSeek 1.3b.

    Source: §5 · p. 37

  • Takeaways & Limitations

    The review identifies potential for agentic workflows to offer greater adaptability, enhanced reasoning, and more consistent educational performance.

    Source: §6 · p. 38

    Educational tool-use systems face interoperability and student-data privacy challenges, alongside bias and educator training needs.

    Source: §4.3.4 · pp. 30–31

Abstract

from arXiv · show

Artificial intelligence (AI) has transformed various aspects of education, with large language models (LLMs) driving advancements in automated tutoring, assessment, and content generation. However, conventional LLMs are constrained by their reliance on static training data, limited adaptability, and lack of reasoning. To address these limitations and foster more sustainable technological practices, AI agents have emerged as a promising new avenue for educational innovation. In this review, we examine agentic workflows in education according to four major paradigms: reflection, planning, tool use, and multi-agent collaboration. We critically analyze the role of AI agents in education through these key design paradigms, exploring their advantages, applications, and challenges. To illustrate the practical potential of agentic systems, we present a proof-of-concept application: a multi-agent framework for automated essay scoring. Preliminary results suggest this agentic approach may offer improved consistency compared to stand-alone LLMs. Our findings highlight the transformative potential of AI agents in educational settings while underscoring the need for further research into their interpretability, trustworthiness, and sustainable impact on pedagogical impact.

1 Introduction

Conventional LLMs in education are constrained by static training data, limited adaptability, and limited nuanced reasoning. The paper reviews four agentic paradigms and presents a multi-agent essay-scoring proof of concept.

Introduction

  • Conventional educational LLM applications rely on static training data and have limited adaptability and nuanced reasoning.The authors connect these constraints to the need for sustainable, responsible AI research.Source: §1 · p. 2
  • Educational agent integration requires careful examination because systems need to be trustworthy and interpretable across learning environments, student populations, and pedagogical methods.The authors call for further research into these settings.Source: §1 · p. 2
  • The paper reviews reflection, planning, tool use, and multi-agent collaboration, analyzing their foundations, educational applications, advantages, and challenges.The review frames these paradigms as ways AI agents may address conventional LLM limitations.Source: §1 · p. 2
  • MASS is a proof-of-concept multi-agent system for automated essay scoring, reported as more consistent and reliable than stand-alone LLMs.The paper presents it to illustrate agentic workflows’ potential in education.Source: §1 · p. 2

2 Background: AI Agents and Agentic Workflows

The paper defines agents as LLM-centered systems that interact with environments toward goals, then outlines agentic workflows and their four recurring paradigms. It also describes the emerging agent layer and platforms for building agentic systems.

2 Background: AI Agents and Agentic Workflows

  • An AI agent uses an LLM as its reasoning engine to control an application and interact with its environment toward specific goals.Agentic workflows structure the operation of autonomous or semi-autonomous agents, which can retrieve current information and use tools.Source: §2 · p. 3
  • Inference-time methods including chain-of-thought, self-consistency, ReAct, and reflection are described as drivers of recent performance improvements amid data and compute constraints.The paper connects this shift to limits in data availability and computational resources (Zeng et al., 2024; Wei et al., 2022; Yao et al., 2023; Shinn et al., 2023).Source: §2 · p. 3
  • Agentic frameworks built on GPT-3.5 and GPT-4 show HumanEval gains, with similar improvements reported for TransCoder and MBPP (Zhong et al., 2024).The text also notes preliminary educational improvements in automated scoring (Guo et al., 2024).Source: §2 · p. 4, Figure 2.2 · p. 4

2.2 AI Agent Platforms

  • The agent layer is presented as orchestrating foundational models, cloud infrastructure, and hardware for complex goal-oriented tasks; platforms vary in structure, flexibility, and architecture support.Named platforms include AutoGen, CrewAI, LangGraph, MetaGPT, and Phidata.Source: §2 · p. 4, §2.2 · p. 6, Table 2.1 · p. 6

2.1 Agentic Workflow

  • Agentic workflows commonly plan and sequence subtasks, retrieve information or use tools, then execute and refine the plan.The degree of agenticness can range from a single-tool system to complex multi-agent collaboration.Source: §2.1 · p. 5
  • The four complementary paradigms are reflection, planning, tool use, and multi-agent collaboration.They form the basis of the paper’s education-focused review.Source: §2.1 · p. 5, Figure 2.4 · p. 6

3 Methodology

The methodology combines a structured literature review with a separately described illustrative case study. It searches and screens education-focused agentic research, analyzes it through four paradigms, and presents findings narratively.

3 Methodology

  • The review synthesizes current research on trends, applications, benefits, and challenges across reflection, planning, tool use, and multi-agent collaboration.Its framework draws on recent surveys and discussions of these paradigms.Source: §3 · p. 7

3.1 Literature Search Strategy

  • Searches covered Google Scholar, Scopus, Web of Science, and ACM Digital Library, using agent, LLM, education, paradigm, and application terms.The search included terms such as ReAct, Reflexion, chain-of-thought, automated tutoring, scoring, and adaptive learning.Source: §3.1 · p. 7

3.2 Selection Criteria and Scope

  • The review focused primarily on publications from 2020 to 2024 and synthesized approximately 93 sources, including peer-reviewed journal articles, conference proceedings, and significant, highly cited arXiv preprints.Source: §3.2 · p. 8
  • Eligible sources addressed agentic mechanisms or architectures, at least one target paradigm, and educational settings or relevant tasks; foundational-LLM-only and unrelated education studies were excluded.The stated scope includes settings such as higher education, K–12, and professional training, and tasks such as assessment, feedback, tutoring, and simulation.Source: §3.2 · p. 8

3.3 Analysis Framework

  • For each paradigm, the analysis extracted conceptual foundations, educational applications, advantages, and challenges or future directions.The framework also considered ethical concerns.Source: §3.3 · pp. 8–9

3.4 Synthesis and Presentation

  • Findings were synthesized narratively by paradigm, with tables summarizing studies, architectures, and concepts; the Section 5 case study used its own implementation and evaluation description.The case study used the dataset from Crossley et al. (2024) and was distinct from the literature-synthesis methodology.Source: §3.4 · p. 9

4 Agentic Paradigms in Education

The section reviews four agentic workflow paradigms, covering their foundations, educational roles and applications, and associated challenges.

4 Agentic Paradigms in Education

  • The review organizes reflection, planning, tool use, and multi-agent collaboration by their foundations, educational roles and applications, and challenges.These are the four paradigms established in the paper’s framework.Source: §4 · p. 9

4.1 Reflection Systems

Reflection systems use iterative assessment and feedback to refine agent behavior, with educational applications spanning tutoring, assessment, and collaboration. Their potential is constrained by computational demands and concerns about privacy, bias, generalization, transparency, and human oversight.

4.1.1 Mechanisms of Reflection Systems

  • Reflection mechanisms combine interactive feedback, verbal reinforcement, and external verification or self-correction, linking internal evaluation with external tools.Agents assess prior actions, identify errors, and adapt strategies through iterative review.Source: §4.1.1 · p. 10, §4.1 · p. 10

4.1.1.1 Interactive Feedback

  • CRITIC uses a verify-then-correct cycle with search, code, or calculator tools to critique and iteratively refine outputs.It complements internal assessment with external validation.Source: §4.1.1.1 · pp. 10–11

4.1.1.2 Verbal Reinforcement

  • Reflexion converts feedback into actionable episodic memory, helping agents recognize recurring errors and adapt across iterations.The paper describes its use in multi-step reasoning and workflows such as debugging.Source: §4.1.1.2 · p. 11

4.1.1.3 External Verification and Self-Correction

  • SELF-REFINE iterates within one model without external supervision, while DeepSeek is presented as adding advanced reasoning and feedback for educational tasks.SELF-REFINE is described as achieving gains with minimal computational resources, especially in tasks needing incremental improvement.Source: §4.1.1.3 · pp. 11–12

4.1.2 Reflection Systems in Education

  • Educational reflection systems are described for adapting tutoring, refining essay feedback and grading, supporting group participation, and prompting learner metacognition.The review also discusses simulations using AI-generated mentors and evaluators.Source: §4.1.2 · pp. 13–14

4.1.3 Challenges and Future Directions

  • Challenges include resource-intensive reflection that limits scalability, sensitive learner data, feedback bias, and difficulty generalizing across disciplines, cultures, and instructional styles.The review also flags limited transparency and calls for secure, inclusive systems, human oversight, and multimodal feedback.Source: §4.1.3 · p. 16

4.2 Planning Systems

Agentic planning decomposes goals into action sequences that can be executed, monitored, and revised. The section describes planning patterns and educational applications, alongside data privacy and bias concerns.

4.2 Planning Systems

  • Decomposition plans subtasks before execution, whereas interleaved planning adapts decomposition, planning, and execution as new subtasks emerge.ReACT cycles through reasoning, acting, observing, and reflecting; ReWOO uses a planner, worker, and solver, with the worker adapting the plan to new information.Source: §4.2 · pp. 17–18
  • Planning ranges from simple action sequences and goal-directed plans to planners that learn from experience or seek new information.The passage names simple, goal-oriented, learning, and self-improving planners; it refers to five categories but the supplied passages describe only these four.Source: §4.2 · pp. 18–19
  • The planning workflow defines a goal, analyzes data, models actions, generates a plan, refines and simulates it, then executes, monitors, and learns from outcomes.The process can use search algorithms such as depth-first search, breadth-first search, or A*.Source: §4.2 · pp. 19–20, Figure 4.1 · p. 19
  • Planning agents can personalize learning pathways by identifying student knowledge gaps and adjusting activities in real time.The review also describes planning for curricula, institutional decisions, collaboration, resource allocation, and equity and access.Source: §4.2 · pp. 20–22
  • Educational planning systems must protect student data and address algorithmic bias through transparent, accountable models and regular fairness audits.The review identifies privacy, security, and potential reinforcement of existing inequalities as concerns.Source: §4.2 · p. 22

4.3 Tool-use systems

Educational tool-use systems connect agents to external functions and data, then process the returned information to generate task-relevant outputs. The section describes tool categories, educational uses, and mechanisms for selecting and coordinating tools.

4.3 Tool-use systems

  • Tool-use agents can combine external resources and functions to support complex tasks, including tutoring tailored to student performance.Examples include calculators, educational databases, and communication interfaces for personalized feedback and practice recommendations.Source: §4.3 · pp. 22–23
  • Tool-use mechanisms include formatted tool requests, post-processing, iterative reasoning, context management, and, in complex cases, collaboration among specialized agents.Tool selection may use heuristics to prioritize functions relevant to the current task.Source: 4.3.1.Components and Mechanisms of… · pp. 23–24
  • The operational workflow moves from task initiation and tool selection through execution, data processing and reflection, and output generation.The passage gives examples including search and database APIs, Python, Pandas, NumPy, and visualization tools.Source: 4.3.1.Components and Mechanisms of… · p. 26

4.3.2. Type of Tool-Use Systems

  • Tool categories include decision and planning tools, simulation and modeling tools, and interaction and communication tools.The supplied examples include planning and scheduling software, simulation tools, conversational frameworks, and virtual-meeting APIs.Source: §4.3.2 · p. 28, Interaction and Communication · p. 28

Interaction and Communication

  • Monitoring and evaluation tools track engagement and outcomes, while governance tools support fairness and ethical accountability.Examples include analytics, formative assessment, fairness indicators, and ethics frameworks.Source: Interaction and Communication · p. 28

4.3.3. Tool - Use systems in Education

  • In education, coordinated tools can provide real-time feedback and personalized learning paths, while analytics can help teachers track progress and adapt instruction.The review also describes automating grading and attendance and using student performance patterns to recommend resources.Source: §4.3.3 · pp. 29–30

4.6 Tool-use systems in Education

The section surveys prior work on educational tool-use and identifies practical and ethical conditions for implementation. It recommends frameworks that align tool use with data management and pedagogical goals.

4.6 Tool-use systems in Education

  • The literature table includes work on AI policy and governance, strategies addressing inequities, AI-in-education paradigms, and autonomous agents.The listed entries also associate these studies with ethics and fairness, communication and interaction, or simulation and modeling tools.Source: §4.6 · p. 30

4.3.4. Limitations and Challenges

  • Tool-use implementation faces interoperability and data-privacy concerns, possible algorithmic bias, and educator training and support needs.The review notes that weak interoperability can fragment learning experiences and that insufficient educator preparation may limit classroom adoption.Source: §4.3.4 · pp. 30–31
  • The review calls for implementation frameworks that establish data-management guidelines, support stakeholder collaboration, and align tools with pedagogical goals.It frames these conditions as part of implementing tool-use systems to support personalized learning and learner agency.Source: §4.3.4 · p. 31

3.4 Multi Agent Systems

The section presents multi-agent collaboration as a way to partition complex tasks among specialized agents, and reviews its designs, educational applications, and challenges.

3.4 Multi Agent Systems

  • Specialized agents can divide complex tasks and manage context more effectively than a single agent, while exchanging feedback and reintegrating subtasks through dialogue.The section identifies single agents’ difficulty choosing among tools, handling large contexts, and covering multiple specialties.Source: §3.4 · pp. 31–32
  • Modularity, specialization, and explicit interaction controls offer oversight and may reduce hallucinations, while allowing components to be updated or replaced independently.Dedicated agents can focus on particular domains; conversations provide explicit mechanisms for overseeing interactions.Source: §3.4 · p. 32

4.4.1 Multi Agent Architectures

  • Supervisor architectures delegate from a lead agent to subagents, whereas network architectures allow agents to communicate with one another.Source: §4.4.1 · p. 32

3.4.4 Multi Agent Systems in Education

  • Educational applications include adaptive learning, specialized e-learning agents, simulations such as PitchQuest, MEDCO, and SimClass, and multi-agent scoring and feedback systems.Earlier multi-agent education systems predate LLMs; reported scoring work includes one agent generating feedback and another validating and refining it.Source: §3.4.4 · p. 34

4.4.4 Limitations

  • Multi-agent systems can increase token use, processing time, computational cost, and environmental impact; poor orchestration or hallucinating agents can also reduce output quality.Source: §4.4.4 · p. 36
  • Development and coordination are complex, and communication breakdowns, biased decisions, or variable agent interactions can burden systems and undermine learner engagement or instructional fairness.Source: §4.4.4 · p. 36

5 Illustrative Application: Multi-agent scoring system

The paper introduces MASS, a proof-of-concept multi-agent essay grader, and evaluates it against stand-alone LLMs on ASAP 2.0. MASS reports lower MAE and score variability than the compared models.

System design

  • MASS uses two subagents supervised by a lead agent: one assesses essay content and the other checks grammar and mechanics; the lead combines their feedback and requests multiple reports when feedback diverges widely.The system is based on GPT-4o.Source: §5 · p. 36, Figure 5.1 · p. 37

Evaluation

  • The evaluation uses ASAP 2.0, about 17,000 argumentative essays scored from 1 to 6, and compares MASS against stand-alone GPT-4o, DeepSeek 67b, and DeepSeek 1.3b.Source: §5 · p. 37

Results

  • 0.5612 MAE for MASS is lower than GPT-4o (0.6129), DeepSeek 67b (0.7345), and DeepSeek 1.3b (1.6956) on automated essay grading.The evaluation results are summarized in Table 5.1.Source: Table 5.1 · p. 37, §5 · p. 37
  • MASS has a standard deviation of 0.8295, below GPT-4o (0.9555), DeepSeek 67b (1.0102), and DeepSeek 1.3b (1.7666), indicating more consistent scores across essays.Source: §5 · p. 38
  • The authors attribute MASS’s performance to separate content and mechanics assessments followed by synthesis, and say its structured approach reduces over-praise and over-inference errors.They also state that its verification mechanism can lead to more refined and balanced scoring.Source: §5 · p. 38

6 Conclusion

The conclusion synthesizes the review of four agentic design paradigms and the MASS proof of concept. It identifies unresolved deployment challenges and calls for research on transparency, fairness, transfer, and long-term effects.

6 Conclusion

  • The review covers reflection, planning, tool use, and multi-agent collaboration, and concludes that agentic workflows have potential for greater adaptability, reasoning, and consistency in education.The MASS proof of concept demonstrated advantages over stand-alone LLMs.Source: §6 · p. 38
  • Real-time data, iterative feedback, and task decomposition are presented as ways agents address conventional LLMs’ static nature and support educational solutions.The essay-scoring framework highlights possible gains in reliability, consistency, and scalability.Source: §6 · p. 38
  • The authors describe reducing computational redundancy and optimizing resource allocation as opportunities for agentic systems to minimize energy use and resource consumption.Source: §6 · pp. 38–39
  • Unresolved challenges include transparency, fairness, accountability, transfer across learners and pedagogical settings, and ensuring AI supports rather than replaces meaningful educational interaction.Future research also includes longitudinal study of effects on learning, skills, teaching, creativity, self-directed learning, and peer collaboration, alongside attention to under-resourced communities.Source: §6 · p. 39, §6.1 · p. 39

6.1 Future Research

  • Future work should improve interpretability, align systems with pedagogical goals, develop fair-deployment policies, and make agent reasoning transparent to educators and students.Source: §6.1 · p. 39

Statements & Declarations

The authors report no funding or other support and no relevant financial or non-financial interests; they also describe author contributions to the study and manuscript.

Statements & Declarations

  • The authors declare that they received no funds, grants, or other support while preparing the manuscript.Source: Statements & Declarations · p. 46
  • The authors state that they have no relevant financial or non-financial interests to disclose.Source: Statements & Declarations · p. 46
  • All authors contributed to conception and design, and the listed authors performed material preparation, data collection, and analysis.Firuz Kamalov wrote the first draft; all authors commented on prior versions and approved the final manuscript.Source: Statements & Declarations · p. 46

Related papers

Loading 2504.20082v1…