Pith. sign in

REVIEW 2 major objections 6 minor 31 references

AI Coding Agents Should Teach, Not Just Code

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-08 16:31 UTC pith:5I3FBPJF

load-bearing objection Position paper with a clear problem framing and reasonable architecture, but the load-bearing assumption about reasoning trace quality is unaddressed. the 2 major comments →

arxiv 2607.06101 v1 pith:5I3FBPJF submitted 2026-07-07 cs.SE cs.AIcs.CYcs.HC

Agents That Teach: Towards Designing Incidental Learning Back into AI-Assisted Software Development

Rohit Mehra , Samdyuti Suri , Prithviraj K Tagadinamani , Kapil Singi , Vikrant Kaulgud , Adam P. Burden This is my paper
classification cs.SE cs.AIcs.CYcs.HC
keywords incidental learningknowledge debtAI coding agentssoftware engineeringskill atrophymulti-agent systemsdeveloper toolscognitive offloading
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that as developers delegate increasing amounts of coding to AI agents, they lose the incidental learning that comes from effortful problem-solving. This creates 'Knowledge Debt'—a developer-level analogue to Technical Debt where the developer cannot fully understand the changes the agent makes. The authors claim this learning will not return on its own and must be consciously designed back into the development workflow. To operationalize this, they propose SHIELD, a multi-agent system that runs alongside the developer's AI coding agent, observes its reasoning traces (alternatives considered, design rationale, uncertainty), and surfaces targeted, non-interruptive microlearning moments calibrated to the individual developer's existing knowledge.

Core claim

The central mechanism is the extraction of pedagogical signals from the AI coding agent's reasoning trace—information about why a design decision was made, what alternatives were considered, and where the agent was uncertain. The paper claims this information cannot be reconstructed from code artifacts alone and that a separate multi-agent system can triage these signals against a per-developer knowledge model to identify genuine knowledge gaps and deliver contextual microlearning without disrupting developer flow.

What carries the argument

SHIELD (Safeguarding Human Expertise and Incidental Learning in Software Development): a multi-agent system operating within the IDE. Key components include the Telemetry Observer Agent (captures the coding agent's reasoning), Teachability Triage Agent (filters for genuine teachable moments using complexity, novelty, and transferability thresholds), Developer's Evolving Concept Map (a graph-based per-developer knowledge model), Probe Generator Agent (verifies whether a gap is real), Knowledge Assessor Agent (evaluates developer responses and updates the Concept Map), and Microlearning Generator Agent (produces calibrated learning content).

Load-bearing premise

The load-bearing assumption is that an AI coding agent's reasoning trace contains reliable, pedagogically meaningful signals—genuine design rationale, real alternatives considered, and honest uncertainty—that a separate LLM-based triage agent can correctly identify as genuine knowledge gaps for a specific developer. There is no evidence presented that these reasoning traces accurately reflect meaningful design thinking rather than post-hoc rationalization, nor that the triage

What would settle it

If developers who use SHIELD show no measurable improvement in independent comprehension or capability compared to those who use standard AI coding agents (as measured by post-task assessments without AI assistance), the core claim that this system repays Knowledge Debt would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reasoning-trace-to-microlearning pipeline works, it establishes a new design category where AI tools are evaluated not just on productivity but on their capacity to sustain and grow developer expertise.
  • Knowledge Debt could become a measurable, trackable metric within software engineering, analogous to how Technical Debt is currently monitored.
  • The separation of the coding agent from the teaching agent suggests a broader architectural principle: models optimized for code generation may be suboptimal for pedagogy, and future AI ecosystems may require specialized role-differentiated agents.
  • If the Concept Map approach proves effective, it could generalize to other domains where AI assistance risks skill atrophy, such as medicine, aviation, or writing.
  • The 'out-of-band' delivery model (asynchronous queues and feeds rather than blocking prompts) could become a standard interaction pattern for any system that needs to inject information into a focused workflow without breaking flow state.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The system's value hinges entirely on the signal-to-noise ratio of the coding agent's reasoning trace. If LLM reasoning traces are post-hoc rationalizations rather than genuine reflections of design rationale, the triage agent would be surfacing fabricated pedagogical content.
  • The Concept Map's cold-start initialization (analyzing developer-authored code to infer familiarity) may systematically misrepresent knowledge for developers who have already been heavily using AI agents, as their authored code may already reflect AI-generated patterns they do not fully understand.
  • If effective, this architecture implies that the optimal AI coding workflow may require agents to deliberately reason more verbosely than necessary for code generation alone, trading some generation efficiency for richer pedagogical signal extraction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This paper introduces the concept of 'Knowledge Debt'—a developer-level analogue to Technical Debt arising from over-reliance on AI coding agents—and proposes six design principles for embedding incidental learning back into developer-agent interactions. The authors present SHIELD, a multi-agent system that instruments an AI coding agent to emit reasoning telemetry, triages that telemetry for teachable moments using a per-developer Concept Map, and delivers asynchronous microlearning through out-of-band IDE channels. The paper is positioned as a vision/NIER contribution and explicitly states that empirical evaluation is future work.

Significance. The paper addresses a timely and important concern: the erosion of developer expertise as coding tasks are increasingly delegated to AI agents. The framing of this as 'Knowledge Debt' is a useful conceptual contribution that is likely to resonate with the software engineering community. The six design principles (Contextual, Grounded, Ambient, Selective, Adaptive, Closed-Loop) are clearly articulated and provide a reasonable scaffold for future system design. The SHIELD architecture is well-motivated and the walkthrough in Figure 2 effectively illustrates the end-to-end pipeline. The distinction between in-band cognitive forcing functions (which interrupt flow) and out-of-band ambient learning (which respects flow) is a valuable design insight. The paper is honest about its status as an early-stage vision with no empirical validation, which is appropriate for the NIER track.

major comments (2)
  1. §3, 'Grounded' principle and §4.1, Telemetry Observer Agent: The paper's central architectural claim depends on AI coding agents emitting structured reasoning telemetry (alternatives considered, design rationale, confidence). The 'Grounded' principle asserts that this reasoning 'carries information that cannot be reconstructed from artifacts alone,' and Figure 2(c) shows a neatly structured telemetry record with fields for reasoning, alternatives, and confidence. However, the paper provides no evidence that current AI coding agents (Claude Code is named in §4.2) actually emit reasoning traces at this level of structure, nor does it address the well-documented concern that LLM chain-of-thought outputs can be post-hoc rationalizations rather than faithful records of deliberation. If the input telemetry does not reliably contain genuine alternatives-considered and uncertainty signals, the T
  2. §4.1, Teachability Triage Agent: The triage agent uses 'configurable thresholds' for complexity, novelty, and transferability to distinguish genuine teachable moments from routine activity. These thresholds are free parameters, but the paper provides no guidance on how they would be initialized, validated, or tuned. Since the entire system's value proposition depends on achieving a high signal-to-noise ratio (surfacing too frequently 'trains developers to ignore the system,' as the paper itself notes in the 'Selective' principle), the absence of any discussion of how these thresholds would be set in practice is a load-bearing gap. A brief discussion of how the authors plan to approach threshold calibration, or at minimum an acknowledgment that this is an open problem requiring empirical investigation, would strengthen the contribution.
minor comments (6)
  1. §1: The term 'Knowledge Debt' is introduced as a novel contribution, but the related work (§2) discusses 'epistemic debt' [26] and 'comprehension debt' [1] as closely related concepts. The paper should explicitly distinguish Knowledge Debt from these prior formulations to clarify its specific contribution and avoid the appearance of rebranding existing ideas.
  2. §4.1: The Concept Map's cold-start initialization is described as analyzing 'code authored by the developer (not AI-generated).' In practice, distinguishing AI-generated from human-authored code at the line level is non-trivial. A brief acknowledgment of this challenge, or a note on how the prototype handles mixed-authorship files, would be helpful.
  3. Figure 1: The architecture diagram is informative but the agent roles and data flows could be more clearly labeled. In particular, the relationship between the Learning Orchestrator and the specialized agents is described in text but not visually distinguished in the figure.
  4. §4.2: The choice of GPT-5.1 as the underlying model for all SHIELD agents is stated without justification. Given the paper's own observation that 'models optimized for code generation may not be well-suited for teaching,' a brief note on why GPT-5.1 was selected over models specifically fine-tuned for instruction or tutoring would strengthen the implementation discussion.
  5. §5: 'Early demonstrations of SHIELD to stakeholders within our organization have generated promising feedback' is vague. For a NIER paper, even a brief characterization of the feedback (e.g., number of stakeholders, types of concerns raised) would add credibility. If no specifics can be shared, this sentence could be removed.
  6. References [1] and [26] are both dated 2026 and appear to be very recent arXiv preprints. The authors should verify these citations are accurate and stable, as the field is moving rapidly and these works may be revised or renamed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for a constructive and encouraging review. The referee correctly identifies the paper's status as an early-stage vision (NIER track) and raises two substantive concerns: (1) whether current AI coding agents actually emit structured reasoning telemetry of sufficient fidelity, and whether LLM chain-of-thought can be trusted as genuine deliberation rather than post-hoc rationalization; and (2) how the configurable thresholds in the Teachability Triage Agent would be initialized, validated, and tuned. Both points are well-taken. We will address them through targeted revisions that acknowledge the telemetry fidelity concern more explicitly, discuss the faithfulness literature, and add discussion of threshold calibration as an open empirical problem requiring investigation.

read point-by-point responses

  1. Referee: §3, 'Grounded' principle and §4.1, Telemetry Observer Agent: The paper's central architectural claim depends on AI coding agents emitting structured reasoning telemetry (alternatives considered, design rationale, confidence). The 'Grounded' principle asserts that this reasoning 'carries information that cannot be reconstructed from artifacts alone,' and Figure 2(c) shows a neatly structured telemetry record with fields for reasoning, alternatives, and confidence. However, the paper provides no evidence that current AI coding agents (Claude Code is named in §4.2) actually emit reasoning traces at this level of structure, nor does it address the well-documented concern that LLM chain-of-thought outputs can be post-hoc rationalizations rather than faithful records of deliberation. If the input telemetry does not reliably contain genuine alternatives-considered and uncertainty signals, the T

    Authors: The referee raises a valid and important concern. We agree that the paper should be more explicit about the current state of AI coding agent telemetry and the faithfulness of chain-of-thought reasoning. We will make the following revisions: (1) In §4.2, we will clarify that our prototype instruments Claude Code through its configuration file to emit structured telemetry, and that the level of structure shown in Figure 2(c) is achieved through this instrumentation rather than being natively available in all AI coding agents. We will note that not all current agents emit reasoning at this level of structure by default, and that SHIELD's approach assumes (or requires) agent cooperation in emitting such telemetry. (2) In §3 (Grounded principle) and §4.1, we will add a discussion of the chain-of-thought faithfulness concern, citing the relevant literature on post-hoc rationalization in LLM reasoning traces. We will acknowledge that the Grounded principle depends on the assumption that agent reasoning traces contain genuine deliberation signals, and that the degree to which this holds is an empirical question that varies across models and configurations. (3) We will note that even where reasoning traces are imperfect or partially post-hoc, they may still carry more teachable signal than artifacts alone, but we agree this is a limitation that should be surfaced rather than assumed away. We believe these revisions honestly address the concern without overclaiming the current state of the art. revision: yes

  2. Referee: §4.1, Teachability Triage Agent: The triage agent uses 'configurable thresholds' for complexity, novelty, and transferability to distinguish genuine teachable moments from routine activity. These thresholds are free parameters, but the paper provides no guidance on how they would be initialized, validated, or tuned. Since the entire system's value proposition depends on achieving a high signal-to-noise ratio (surfacing too frequently 'trains developers to ignore the system,' as the paper itself notes in the 'Selective' principle), the absence of any discussion of how these thresholds would be set in practice is a load-bearing gap. A brief discussion of how the authors plan to approach threshold calibration, or at minimum an acknowledgment that this is an open problem requiring empirical investigation, would strengthen the contribution.

    Authors: The referee is correct that threshold calibration is a load-bearing gap in the current paper. We will add a discussion in §4.1 that addresses this directly. Specifically, we will: (1) Acknowledge that the thresholds for complexity, novelty, and transferability are free parameters whose optimal values are unknown and cannot be set a priori without empirical investigation. (2) Outline our planned approach to calibration: starting with conservative (high) thresholds to minimize false positives and avoid training developers to ignore the system, then progressively lowering them based on developer engagement metrics (e.g., probe response rates, microlearning completion rates, comprehension check outcomes) gathered during empirical studies. (3) Note that the Closed-Loop principle already provides a mechanism for this: comprehension check outcomes feed back into the Concept Map and can inform threshold adjustment over time, though we agree this is not currently discussed as a calibration mechanism. (4) Explicitly state that threshold calibration is an open problem requiring empirical investigation, which we plan to address in the user studies described in §5. We believe this is an honest acknowledgment that the problem is real and that the paper's contribution is in framing the problem and proposing an architecture, not in claiming that the thresholds are solved. revision: yes

Circularity Check

0 steps flagged

No circularity: position/architecture paper with no fitted parameters, no equations, and no self-citation chain.

full rationale

This is a NIER-track position paper proposing six design principles and a multi-agent architecture (SHIELD). There are no equations, no fitted parameters, no mathematical derivations, and no predictions that reduce to inputs by construction. The design principles (§3) are stated independently of SHIELD's implementation (§4), and SHIELD is presented as one possible instantiation rather than a derivation from the principles. The references are all to external work (cognitive offloading studies, skill atrophy literature, AI coding tools); there are no self-citations by the paper's own authors. The 'Grounded' principle's assumption that agent reasoning traces contain pedagogically meaningful signals is an empirical assumption about the world, not a circular definition—it can be falsified independently. Whether that assumption is correct is a matter of correctness risk, not circularity. The paper is self-contained as a position paper and exhibits no circular reasoning patterns from the enumerated list.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 3 invented entities

The axiom ledger reflects that this is a position/architecture paper with no empirical validation. The free parameters are architectural configuration choices with no specified values. The axioms are domain assumptions about the pedagogical quality of LLM reasoning traces and developer engagement with voluntary learning content. The invented entities (Knowledge Debt, Concept Map, SHIELD) are conceptual contributions without independent falsifiable handles.

free parameters (2)
  • Teachability signal thresholds (complexity, novelty, transferability) = Not specified; described as 'configurable'
    §4.1 states the Teachability Triage Agent evaluates concepts against 'a set of configurable teachability signals, including complexity, novelty, and transferability, each governed by configurable thresholds.' No values, tuning methodology, or validation of threshold choices is provided.
  • Concept Map initialization parameters = Not specified
    §4.1 states the Concept Map 'can be initialized by analyzing the code authored by the developer... under the assumption that concepts present in that code are already familiar.' The analysis method and familiarity criteria are unspecified.
axioms (4)
  • domain assumption AI coding agent reasoning traces contain reliable pedagogical signals that cannot be reconstructed from code artifacts alone.
    §3 (Grounded principle) and §4.1 assert that the agent's reasoning carries information about design rationale, alternatives, and uncertainty that distinguishes genuine teachable moments from routine activity. This is the foundational assumption for the entire architecture and is untested in the paper.
  • domain assumption Developers will engage with out-of-band learning content voluntarily.
    §3 (Ambient principle) and §4.1 assume developers will choose to engage with the Probe Queue and Microlearning Feed. No evidence is presented that developers will actually use these channels in practice, especially under productivity pressure.
  • domain assumption LLM-generated probes and comprehension checks accurately assess developer knowledge gaps.
    §4.1 describes the Probe Generator Agent and Knowledge Assessor Agent generating and evaluating probes. The accuracy of LLM-based knowledge assessment is assumed but not validated.
  • domain assumption Skill atrophy from AI coding agents is a real and measurable phenomenon.
    §1 cites supporting evidence (Shen & Tamkin [27]: 17% lower comprehension scores with AI assistance), but the paper's specific framing of 'Knowledge Debt' as a compounding, measurable construct is asserted, not demonstrated. The paper itself lists operationalizing Knowledge Debt as future work.
invented entities (3)
  • Knowledge Debt no independent evidence
    purpose: Framing concept for the accumulating gap between AI-generated code changes and developer understanding, positioned as a developer-level analogue to Technical Debt.
    Introduced in §1 as a new term. The paper acknowledges closely related prior concepts (epistemic debt [26], comprehension debt [1]) in §2. No independent falsifiable handle is provided; the paper lists operationalizing Knowledge Debt as a measurable construct as future work.
  • Developer's Evolving Concept Map no independent evidence
    purpose: Per-developer graph representation of concepts the developer has demonstrated familiarity with, been taught, or assessed on, used to calibrate learning interventions.
    Described in §4.1 as a Neo4j graph. No evidence is presented that this representation accurately models developer knowledge or that it improves intervention quality compared to simpler approaches.
  • SHIELD multi-agent system no independent evidence
    purpose: Operational instantiation of the six design principles, comprising Telemetry Observer, Learning Orchestrator, Teachability Triage, Probe Generator, Knowledge Assessor, and Microlearning Generator agents.
    Described in §4 with a prototype implementation, but no empirical evaluation of the system's effectiveness is presented. The paper states 'Early demonstrations of SHIELD to stakeholders within our organization have generated promising feedback' (§5), which is not independent evidence.

pith-pipeline@v1.1.0-glm · 11950 in / 3337 out tokens · 574519 ms · 2026-07-08T16:31:04.022589+00:00 · methodology

0 comments
read the original abstract

AI coding agents are rapidly reshaping how software is built, with developers increasingly delegating substantial coding tasks to autonomous agents in pursuit of higher productivity. While these gains are real, they come at the cost of incidental learning. Developers historically acquired informal knowledge through effortful problem-solving, and this has long shaped how software engineering expertise develops. However, with over-reliance on agentic coding, unpracticed skills could atrophy silently over time. As this learning pathway is short-circuited, developers risk silently accruing Knowledge Debt, a developer-level analogue of Technical Debt, where changes the agent executes that the developer cannot fully understand accrue over time. In this paper, we argue that incidental learning will not re-emerge on its own and must be consciously designed back into developer-agent interactions, and propose six design principles to guide such systems. We then present "SHIELD", a multi-agent system grounded in the notion of "agents that teach", that operationalizes these principles by leveraging the AI coding agent's own reasoning to surface contextual, out-of-band learning moments without disrupting developer flow. Through this work, we envision a path toward learning-aware development environments where productivity and learning are complementary, not competing.

Figures

Figures reproduced from arXiv: 2607.06101 by Adam P. Burden, Kapil Singi, Prithviraj K Tagadinamani, Rohit Mehra, Samdyuti Suri, Vikrant Kaulgud.

Figure 1
Figure 1. Figure 1: SHIELD architecture: A multi-agent system for em￾bedding incidental learning in developer-agent interactions. • Closed-Loop: The system must verify whether the learning has actually been internalized, through lightweight probes or comprehension checks, rather than treating delivery as success. Outcomes from each intervention should be used to update the system’s understanding of what to surface, deprioriti… view at source ↗
Figure 2
Figure 2. Figure 2: SHIELD walkthrough using an example of a payment API webhook retry task. through the Microlearning Feed, where it remains available for the developer to engage with at a time of their choosing. While the mi￾crolearning can be delivered as passive content for the developer to consume, the architecture can also accommodate emerging agentic tutors that more actively and effectively impart the same learning th… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 3 internal anchors

  1. [1]

    Muhammad Ovais Ahmad. 2026. Comprehension Debt in GenAI-Assisted Soft- ware Engineering Projects. arXiv:2604.13277 [cs.SE] https://arxiv.org/abs/2604. 13277

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Hewa Fouad Ali, Lisa Jamal Nakshbandi, Fatima Saadi, and Sami Hussein Hakeem Barzani. 2024. The Effect of Spell-Checker Features on Spelling Competence among EFL Learners: An Empirical Study.International Journal of Social Sciences and Educational Studies9, 3 (Sep. 2024), 101–111. doi:10.23918/ijsses.v9i3p101

    work page doi:10.23918/ijsses.v9i3p101 2024

  3. [3]

    Anthropic. 2026. Claude Code. https://claude.com/product/claude-code Ac- cessed: 2026-05-02

    work page 2026

  4. [4]

    Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos. 2021. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI- assisted Decision-making.Proc. ACM Hum.-Comput. Interact.5, CSCW1, Article 188 (April 2021), 21 pages. doi:10.1145/3449287

    work page internal anchor Pith review doi:10.1145/3449287 2021

  5. [5]

    Hung-Fu Chang, MohammadShokrolah Shirazi, Lizhou Cao, and Supannika Kool- manojwong Mobasser. 2025. Coding With AI: From a Reflection on Industrial Practices to Future Computer Science and Software Engineering Education. arXiv:2512.23982 [cs.SE] https://arxiv.org/abs/2512.23982

    work page arXiv 2025

  6. [6]

    CrewAI. 2026. CrewAI: Fast and Flexible Multi-Agent Automation Framework. https://github.com/crewAIInc/crewAI Accessed: 2026-05-02

    work page 2026

  7. [7]

    Zheyuan Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz. 2025. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. SSRN Working Paper No. 4945566. https://ssrn.com/abstract=4945566

    work page 2025

  8. [8]

    Cursor. 2026. Cursor. https://www.cursor.com Accessed: 2026-05-02

    work page 2026

  9. [9]

    Louisa Dahmani and Véronique D. Bohbot. 2020. Habitual use of GPS negatively impacts spatial memory during self-guided navigation.Scientific Reports10, 1 (2020), 6310. doi:10.1038/s41598-020-62877-0

    work page doi:10.1038/s41598-020-62877-0 2020

  10. [10]

    Gajos and Lena Mamykina

    Krzysztof Z. Gajos and Lena Mamykina. 2022. Do People Engage Cognitively with AI? Impact of AI Assistance on Incidental Learning. InProceedings of the 27th International Conference on Intelligent User Interfaces(Helsinki, Finland) (IUI ’22). Association for Computing Machinery, New York, NY, USA, 794–806. doi:10.1145/3490099.3511138

    work page doi:10.1145/3490099.3511138 2022

  11. [11]

    Michael Gerlich. 2025. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking.Societies15, 1 (2025). doi:10.3390/soc15010006

    work page doi:10.3390/soc15010006 2025

  12. [12]

    GitHub. 2026. GitHub Copilot. https://github.com/features/copilot Accessed: 2026-05-02

    work page 2026

  13. [13]

    Chen Gong and Yang Yang. 2024. Google effects on memory: a meta-analytical review of the media effects of intensive Internet search behavior.Frontiers in Public HealthVolume 12 - 2024 (2024). doi:10.3389/fpubh.2024.1332030

    work page doi:10.3389/fpubh.2024.1332030 2024

  14. [14]

    Philipp Hugenroth, Valdemar Danry, and Pattie Maes. 2026. Critical Inker: Scaf- folding Critical Thinking in AI-Assisted Writing Through Socratic Questioning. arXiv:2604.07167 [cs.HC] https://arxiv.org/abs/2604.07167

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Majeed Kazemitabaar, Oliver Huang, Sangho Suh, Austin Z Henley, and Tovi Grossman. 2025. Exploring the Design Space of Cognitive Engagement Tech- niques with AI-Generated Code for Enhanced Learning. InProceedings of the 30th International Conference on Intelligent User Interfaces (IUI ’25). Association for Computing Machinery, New York, NY, USA, 695–714. ...

    work page doi:10.1145/3708359 2025

  16. [16]

    Georgios Kostopoulos, Vasileios Gkamas, Maria Rigou, and Sotiris Kotsiantis

    work page

  17. [17]

    doi:10.1109/ACCESS.2025.3620473

    Agentic AI in Education: State of the Art and Future Directions.IEEE Access13 (2025), 177467–177491. doi:10.1109/ACCESS.2025.3620473

    work page doi:10.1109/access.2025.3620473 2025

  18. [18]

    Nord, and Ipek Ozkaya

    Philippe Kruchten, Robert L. Nord, and Ipek Ozkaya. 2012. Technical Debt: From Metaphor to Theory and Practice.IEEE Software29, 6 (2012), 18–21. doi:10.1109/ MS.2012.167

    work page 2012

  19. [19]

    Macnamara, Ibrahim Berber, M

    Brooke N. Macnamara, Ibrahim Berber, M. Cenk Çavuşoğlu, Elizabeth A. Krupin- ski, Naren Nallapareddy, Nicholas E. Nelson, Philip J. Smith, Amy L. Wilson- Delfosse, and Soumya Ray. 2024. Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers’ aware- ness?Cognitive Research: Principles and Implic...

    work page doi:10.1186/s41235-024-00572-8 2024

  20. [20]

    1990.Informal and incidental learning in the workplace

    Victoria J Marsick and Karen E Watkins. 1990.Informal and incidental learning in the workplace. Routledge, London

    work page 1990

  21. [21]

    Microsoft. 2026. Microsoft Azure Cloud Computing Platform. https://azure. microsoft.com Accessed: 2026-05-02

    work page 2026

  22. [22]

    Microsoft. 2026. Visual Studio Code. https://code.visualstudio.com Accessed: 2026-05-02

    work page 2026

  23. [23]

    Neo4j. 2026. Neo4j Graph Database. https://neo4j.com/product/neo4j-graph- database/ Accessed: 2026-05-02

    work page 2026

  24. [24]

    Merle, Andreas Pollreisz, Siegfried K

    Ariel Yuhan Ong, David A. Merle, Andreas Pollreisz, Siegfried K. Wagner, Mertcan Sevgi, Pearse A. Keane, Roman Huemer, Julian Oehling, Markus Jäger, and Josef Huemer. 2026. Flight rules for clinical AI: lessons from aviation for human-AI collaboration in medicine.npj Digital Medicine9, 1 (31 jan 2026), 201. doi:10. 1038/s41746-026-02410-1

    work page 2026

  25. [25]

    OpenAI. 2026. GPT-5.1. https://developers.openai.com/api/docs/models/gpt-5.1 Accessed: 2026-05-02

    work page 2026

  26. [26]

    Elise Paradis, Kate Grey, Quinn Madison, Daye Nam, Andrew Macvean, Vahid Meimand, Nan Zhang, Ben Ferrari-Church, and Satish Chandra. 2025. How Much Does AI Impact Development Speed? an Enterprise-Based Randomized Controlled Trial. In2025 IEEE/ACM 47th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP)(Ottawa, ON...

    work page doi:10.1109/icse-seip66354.2025.00060 2025

  27. [27]

    Epistemic Debt

    Sreecharan Sankaranarayanan. 2026. Mitigating "Epistemic Debt" in Gen- erative AI-Scaffolded Novice Programming using Metacognitive Scripts. arXiv:2602.20206 [cs.SE] https://arxiv.org/abs/2602.20206

    work page arXiv 2026

  28. [28]

    Judy Hanwen Shen and Alex Tamkin. 2026. How AI Impacts Skill Formation. arXiv:2601.20245 [cs.CY] https://arxiv.org/abs/2601.20245

    work page arXiv 2026

  29. [29]

    Sonar. 2026. State of Code Developer Survey Report: The Current Reality of AI Coding. Sonar Blog. https://www.sonarsource.com/blog/state-of-code- developer-survey-report-the-current-reality-of-ai-coding/ Accessed: 2026-05- 02

    work page 2026

  30. [30]

    Stiffler, and Thuy Ngoc Nguyen

    Ba-Thinh Tran-Le, Patrick Thomas, Nicholas M. Stiffler, and Thuy Ngoc Nguyen

    work page

  31. [31]

    2026), 40880–40888

    How Does LLM-powered Coding Assistance Shape Incidental Learning? Exploring Cognitive Forcing Strategies in Programming Education.Proceedings of the AAAI Conference on Artificial Intelligence40, 48 (Mar. 2026), 40880–40888. doi:10.1609/aaai.v40i48.42121

    work page doi:10.1609/aaai.v40i48.42121 2026