AGENTICS 2026 Abstracts

Area 1 - Risks and Ethical Concerns

Full Papers

Paper Nr:	46
Title:	Automating Accountable Risk Discovery in Mission-Critical SaaS: Detecting Systemic Interface Gaps via Role-Based Agents
Authors:	Arisa Morozumi, Kazuhiko Tsuda and Hisashi Hayashi
Abstract:	Mission-critical Software-as-a-Service (SaaS) environments involve complex interactions among diverse stakeholders. While conventional risk management focuses primarily on individual software defects, major hazards often emerge at the interfaces between these stakeholders. Centralized audits frequently overlook these interface risks because they rely on a single, all-knowing auditor model. Having unrestricted access to the complete system context creates an "omniscience bias," preventing the audit from reproducing the localized information failures that human actors experience in reality. Consequently, under Normal Accident Theory, these unaddressed "Boundary Conflicts" can trigger systemic failure even when individual actors maintain localized rule compliance. To address this limitation, we introduce Agent-based Multi-perspective Auditing (AMPA), a formalized framework that utilizes role-based Large Language Model (LLM) agents as proxies to model the Bounded Rationality of human stakeholders. By assigning role-specific constraint variables to restrict each agent’s observation space, the AMPA framework systematically uncovers socio-technical vulnerabilities typically discarded as noise. Validation using a real-world food allergy management platform, ALLESC, suggests that our approach consistently maps identified risks to established public safety guidelines. Furthermore, the framework uncovers a "Security Paradox" where rigid compliance with security rules inadvertently induces new physical and ethical hazards. This architectural approach facilitates Shift-Left governance by identifying precursors of systemic negligence during the requirement definition step, establishing a practical mechanism for proactive accountability.

Paper Nr:	145
Title:	ARISMA: Guidelines for AI- and LLM-Assisted Systematic Reviews, Scoping Reviews, and Mapping Studies
Authors:	Mahyar Tourchi Moghaddam and Mina Alipour
Abstract:	Systematic reviews, scoping reviews, mapping studies, and related evidence syntheses are increasingly difficult to conduct with fully manual workflows as search volumes, update cycles, and synthesis requirements continue to expand. At the same time, artificial intelligence, machine learning, and large language models are rapidly entering review practice across query formulation, screening, extraction, categorization, appraisal support, and reporting. Yet the empirical evidence remains uneven, task-dependent, and insufficient to justify unconstrained automation. Existing standards such as PRISMA 2020, PRISMA-S, PRISMAScR, PRISMA-P, PRESS, and SWiM remain essential, but none provides an end-to-end operational standard for when AI use is methodologically appropriate, how it should be validated, which review decisions must remain human-led, and how AI involvement should be reported so that readers can audit it. This paper proposes ARISMA, an AI Reporting and Integration standard for Systematic Methods and Analysis. ARISMA treats AI as an inspected, benchmarked, logged, and reversible assistant rather than an autonomous reviewer. It is built around one governing principle: every consequential scientific decision must remain humaninterpretable, human-auditable, and human-accountable. The paper contributes a lifecycle taxonomy, process guidance, stepwise recommendations across the review pipeline, a governance and provenance model, a tool-support framework, an AI-integrated reporting checklist, and a validation matrix. It also addresses legal, privacy, infrastructure, and sustainability considerations. The framework was iteratively refined through structured expert consultation. The result is a practical and auditable guideline for responsible AI-assisted evidence synthesis.

Paper Nr:	158
Title:	A Survey of Decentralized LLM-Based Multi-Agent Collaboration for Agentic AI
Authors:	Yazan Andoni, Ali Basta, Awid Vaziry, Thomas Cory, Wolf Rieder and Axel Küpper
Abstract:	Current large language models (LLMs) support multi-agent systems (MAS) that reason, plan, and allocate tasks according to roles. In practice, most of these systems run in a centralized environment where all agents operate within the same framework, using predefined workflows, centralized management, and an underlying assumption of mutual trust. This assumption breaks down when the agents are created and managed by different entities and need to collaborate cross-organizationally by locating each other, communicating and assigning tasks. This paper reviews decentralized LLM-based MAS systematically, covering systems in which agents may differ in ownership, trust domain, information access, capabilities, or protocol environment. From 2742 records across five databases, LLM-assisted relevance scoring and four screening stages retained 29 papers. One pattern runs through them: agents connect more easily than they trust each other. Interoperability protocols currently address discovery, message passing, and access to tools; however, they fail to resolve issues of behavioral trustworthiness, coordination semantics, incentive compatibility, multi-hop delegation and accountability. Recent benchmark evaluations suggest that LLM-based MAS can exhibit failure modes such as hidden-information failure, collective false memories, adversarial manipulation, and topology-dependent failures; however, very few benchmarks involve open environments. These observations form the basis of the identified five research gaps for decentralized LLM-based MAS.

Short Papers

Paper Nr:	160
Title:	When the Buyer Is an AI Agent: A Consumer-Protection Risk Taxonomy and Rubric for Agentic Commerce
Authors:	Khalil Israfilzade
Abstract:	AI shopping agents are increasingly able to evaluate options and complete purchases on behalf of consumers. As this delegation deepens, commercial influence begins to target the consumer's agent rather than the consumer, while consumer protection, both psychological and regulatory, continues to assume a consumer who can see the influence attempt. This conceptual paper introduces delegated vulnerability: the consumer's exposure to influence attempts that target the delegated agent, bypass the consumer's persuasion-coping capacities, and inherit the agent's systematic and scalable susceptibilities. Building on the persuasion knowledge model and consumer vulnerability theory, the paper develops a taxonomy of agent-directed influence that crosses three sources (merchants, platforms, and agent providers) with four mechanisms (machine-facing content optimisation, covert content, architectural steering, and preference-layer exploitation), grades the available evidence for each of the twelve resulting cells, and proposes a transparency-counterfactual test that separates legitimate optimisation from manipulation. A four-criterion risk rubric — detectability, attributability, reversibility, and scale — is then applied to three scenarios and reverses the intuitive ordering of severity: acute fraud proves comparatively governable, whereas gradual preference-profile drift approaches a profile in which every existing corrective mechanism fails at once. Five propositions translate the framework into testable claims, and the paper closes with theoretical, managerial, and policy implications, arguing that protection in agentic commerce must come mainly from the supply side.

Area 2 - Models

Full Papers

Paper Nr:	28
Title:	Hallucination Propagation in Multi-Agent LLM Systems: Survival, Distance, Amplification, and Mutation
Authors:	Samir Jabbar, Christophe Cruz and Maria Alice Bertolim
Abstract:	Multi-agent systems built on large language models (LLMs) are increasingly deployed for complex reasoning tasks, yet hallucination behavior in this agentic setting remains markedly underexplored: while hallucinations have been studied extensively in single-model settings, how they survive, propagate, and amplify across interacting agents has received little systematic attention. We present a controlled study com- paring five architectural families single-agent ReAct (F1), state-graph su- pervision (F2), hierarchical delegation (F3), round-robin consensus (F4), and autonomous handoff swarm (F5) across 210 tasks spanning two datasets (HotpotQA, GSM8K 4.1), two injection modes (synthetic and HaluEval sourced), and three random seeds. We introduce a two-layer metric framework: Layer 1 adapts established metrics (faithfulness de- composition, task success, hallucination rate with bootstrap confidence intervals), while Layer 2 contributes novel propagation-specific measures (survival rate, propagation distance, amplification factor, mutation rate). Inter-rater reliability is validated via dual-judge evaluation (Cohen’s κ = 0.43 3.5). Results show that the state-graph architecture (F2) achieves the highest resilience, lowest hallucination survival (14.3%), highest task success (52.8%), and minimal propagation distance (0.41) while the con- sensus based architecture (F4) is actively harmful: it amplifies injected hallucinations (survival 76.5%, propagation distance 7.52) and collapses task success to 10%. An applied study on 450 cardiology multiple-choice questions (MCQs 6) confirms that family rankings transfer to a safety- critical medical domain: the state-graph architecture achieves the lowest hallucination survival score, and dual-judge agreement rises to κ = 0.80. These findings demonstrate that architectural choice is a first-order de- terminant of hallucination resilience in multi-agent LLM systems, robust across task domains.

Paper Nr:	133
Title:	Sponsored Discovery Mechanism for the Web of Open Agents
Authors:	Filipe L. Benevides, José Ernesto Stelzer Monar and Julio Cesar dos Reis
Abstract:	As autonomous LLM agents become primary consumers of web services, they discover and delegate to one another in open ecosystems — a Web of Open Agents — yet existing discovery infrastructures offer no principled way to govern how agents compete for visibility, risking pay-to-win dynamics in which advertising spend, not merit, drives selection. We propose Sponsored Discovery, whose core is macs (Matchgated Agent Composite Score): semantic domain match acts as a multiplicative gate that structurally bounds out-of-domain agents regardless of bid or reputation, coupled with a multi-dimensional behavioral reputation over accuracy, latency, and cost. Using real LLM agents answering MMLU-Pro questions across three adversarial scenarios (honest baseline, domain-falsifying fraudster, SLA/cost-inflated agent), macs attains the highest delegation success in all three; a paired ablation shows the gate’s advantage over a purely additive composite is decisive under domain misreporting and marginal in benign settings, where behavioral reputation alone suffices, while bid-only and quality-adjusted baselines are exploited. We also analyze its incentive properties and limits (calibrated misrepresentation, Sybil whitewashing).

Paper Nr:	135
Title:	A Planner Cache for Short-Term Memory Management in LLM-Based Agents
Authors:	Renan dos Santos Morais, Frances A. Santos, Leandro A. Villas and Julio Cesar dos Reis
Abstract:	LLM-based agents applied to multi-step tasks fundamentally depend on how they organize their short-term memory. In practice, most implementations accumulate the planning state in a linear textual history without explicitly representing the hierarchical relationships among the main objective, subgoals, and dependencies. This study proposes Planner Cache, a tree-based representation for short-term planning memory whose topology is entirely induced by the prompt, with the LLM deciding and emitting the structure at each turn, combined with explicit chain-of-thought reasoning to improve sufficiency evaluation. We compare Planner Cache with three reference configurations in a ReAct agent applied to the HotpotQA benchmark in the distractor setting: an agent without iterative memory, a linear-history-based approach, and a Tree of Thoughts (ToT) configuration with external BFS search. Evaluated using Exact Match (EM) and F1, Planner Cache achieved 70.1% EM and 79.7% F1, outperforming all reference configurations while issuing fewer model calls per episode than any other iterative condition. Ablating each of the three components reduced EM by four to five points. Exposing the full tree rather than only the active branch also increased the context volume per call. Full plan reconstruction caused the largest accuracy drop in episodes that triggered replanning. Finally, removing explicit chain-of-thought reasoning reduced EM by 4.9 points, despite a slight increase in model calls.

Short Papers

Paper Nr:	41
Title:	AASN: Automated Multi-Agent Synthesis and Discovery via Compositional Agent DNA
Authors:	Muntaser Syed, Marius Calin Silaghi, Rwaida Alssadi, Mohammed Alghamdi, Ali Alqarni and Akram Al Ghanmi
Abstract:	Current multi-agent systems require manual construction, lack scalable discovery mechanisms, and fragment across incompatible protocols. We present AASN (Autonomous Agent Synthesis Network), a framework that automatically generates multi-agent systems from natural language task specifications and enables semantic agent discovery at internet scale. AASN introduces three contributions: (1) Agent DNA, a compositional representation encoding agent capabilities as combinable genes with regulatory sequences and mutation operators, enabling synthesis through genetic programming; (2) a three-tier hierarchical discovery architecture supporting semantic capability matching over millions of agents with sublinear query complexity; and (3) a unified protocol adapter providing bidirectional translation across MCP, A2A, and ADK ecosystems with verified semantic preservation. A Task-to-Agent Compiler uses Monte Carlo Tree Search over architectural spaces with learned heuristics to convert natural language specifications into executable multi-agent workflows. An evolution engine enables continuous adaptation through online reinforcement learning and modular skill transfer. We evaluate AASN across seven experiments spanning 50 tasks in 10 domains, demonstrating competitive task success with automated multi-agent generation, discovery at P95 latency under 0.2 ms for one million agents, protocol translation overhead below 0.02% at realistic execution times with perfect semantic preservation across 9,999 roundtrip tests, and positive domain knowledge transfer averaging 10.4% improvement across related fields.

Paper Nr:	89
Title:	Towards Epistemically Grounded Multi-Agent Planning: A Neuro-Symbolic Position
Authors:	Mayar Osama, Nourhan Ehab and Mervat AbuElkheir
Abstract:	Recent advances in Large Language Models (LLMs) have enabled collaborative multi-agent systems (MAS) capable of addressing diverse tasks through coordination and role specialization; however, their use in planning remains fundamentally limited. LLMs are prone to hallucination and offer no formal guarantees, while symbolic MAS suffers from scalability challenges and strong assumptions about complete world models. As a result, neither paradigm adequately addresses coordination under distributed and incomplete knowledge. Despite growing interest in hybrid neuro-symbolic approaches, existing integrations primarily focus on orchestration, memory, and workflow control, while explicit epistemic-planning and belief-revision frameworks from knowledge representation and reasoning (KR) remain largely unexplored. We argue that epistemic logic is not merely a useful extension, but a necessary foundation for reliable LLM-based multi-agent planning, and we outline a principled neuro-symbolic architecture that realizes this vision.

Paper Nr:	98
Title:	An Autonomous LLM-Based Agent Architecture for Cognitive Industrial Diagnostics
Authors:	Emanuel Ribeiro, Tiago Pinto, Arsénio Reis and João Barroso
Abstract:	Predictive machine learning models, such as XGBoost, have significantly advanced industrial maintenance. However, their "black-box" nature remains a major barrier to human trust on the shop floor, as traditional Explainable AI (XAI) outputs like SHapley Additive exPlanations (SHAP) values are often too mathematically abstract for maintenance technicians. This position paper introduces Oráculo, an autonomous agent ecosystem based on Large Language Models (LLMs) designed to transform industrial diagnostics through agentic, conversational interpretation. Instead of merely acting as a text generator, the LLM serves as a cognitive maestro. Through autonomous tool calling, the agent orchestrates predictive models and a K-Nearest Neighbors (KNN) driven Case-Based Reasoning system, which functions as an institutional memory. The ecosystem synthesizes these inputs, translating complex statistical feature importances into clear, natural language diagnostics and actionable maintenance recommendations. We outline the ongoing implementation and prompt design strategies of this approach, discussing preliminary insights and open challenges regarding its future deployment within a real-world industrial setting. Ultimately, this ongoing study establishes a roadmap for trustworthy Artificial Intelligence (AI) centered on human operators in industrial environments, advocating for a paradigm shift where agents do not just predict faults but actively interpret them for human operators.

Paper Nr:	134
Title:	MCPForge: Automatic Generation of Certified Tool-Use Benchmarks for LLM Agents
Authors:	Alejandro Núñez Arroyo and Julio Cesar dos Reis
Abstract:	Large language model (LLM) agents increasingly rely on external tools to perform tasks, yet evaluating their ability to use these tools effectively remains challenging. Because the same answer may be produced from parametric memory, static benchmarks are vulnerable to training-data leakage, manually crafted test cases are costly and limited to predefined tool sets, and LLM-based judges introduce additional computational costs and stochasticity. Evaluations conflate three notions of solvability: whether an executing solution exists, whether the prompt describes it, and whether a model finds it. We present MCPForge, a framework for the automatic and certified generation of tool-use test cases for agents over the Model Context Protocol. The generator composes toolchains over a typed capability graph, executes a gold path against live MCP servers to obtain a deterministic ground truth, derives atomic equality-checkable claims, and scores answers programmatically without an LLM judge. Every case must pass four certification gates (solvable, tool-dependent, distractor-effective, no-leak) against the live agent before admission. On a benchmark of 52 certified cases, certification attains a 95% yield with a size bounded by the substrate’s fact space. Evaluating gpt-oss:20b across three reasoning-effort levels, the benchmark measures tool reliance and identifies where failures occur. Results reveal a retrievalto-synthesis difficulty gradient and a reasoning-effort trade-off: increased reasoning improves synthesis, but degrades single-step retrieval.

Area 3 - Applications

Full Papers

Paper Nr:	39
Title:	Evaluation of Alternative-Based Information Systems for Deliberative Polling Using an Agentic Simulator
Authors:	Rwaida Alssadi, Khulud Alawaji, Balaji Kasula, Muntaser Syed, Badria Sulaiman Alfurhood, Markus Zanker and Marius Silaghi
Abstract:	Deliberative polling promises to improve collective decision-making by exposing shareholders to a broad range of arguments before they vote. Yet ensuring that every voter encounters a representative sample of the reason space, the coverage problem, remains an open challenge, particularly at scale and in adversarial or strategically motivated electorates. This paper introduces a way of evaluating solutions using the LLM-based Agentic Bipolar Argumentation Simulator (ABAS), grounded in a framework which formalises a poll as a six-tuple ⟨Jend, Jopp, Ratt, Renh, VA, VR⟩ of endorsing and opposing justifications, attack and enhance relations, and shareholder- and relation-weights. ABAS simulates N autonomous shareholder agents, each assigned a latent opinion according to desired distributions in [−1, 1], who sequentially vote, choose or author justifications, and optionally submit argumentation-graph links. The simulator implements recommendations that rank existing justifications by their observable endorsement mass. It evaluates the mechanism’s success by coverage, namely the fraction of the corpus reason-tag set represented in the K recommendations presented to each shareholder, as a solution to the NP-hard Subsuming Justification Problem. Reported experiments characterise how creativity rate (pown), recommendation size (K), argumentation density (plinks), and population size (N) affect coverage and corpus diversity. In an authenticated electorate where Sybil attacks are impossible and only the relation graph is gameable, we stress-test the scoring with coordinated strategic voting attacks: a significant tag-flood attack collapses coverage, while author-count relation weighting through a reversed-PageRank rule resists the flood markedly better than uniform weights. ABAS also supports repeated interaction in deliberation, persistent corpus archiving, simulation of strategic opinion presentation, and an interactive browser for auditability.

Paper Nr:	104
Title:	Agentic Accountability: Deterministic Decision Boundaries for High-Stakes Agent Systems
Authors:	Nimrod Foldvari and Florin Craciun
Abstract:	Local churches in the Seventh-day Adventist (SDA) Church collect weekly tithes and offerings on paper: donor slips, cash counting, money forwarded upward through a hierarchy, and an annual regional audit. An earlier phase digitalized this in a deployed serverless application, but the judgment-heavy work stayed manual—transcribing slips, reading bank statements, and judging whether a signed year-end attestation may be accepted. We ask whether LLM agents can take on that work without eroding accountability. Our answer is a design stance we call agentic accountability. The model may only perceive documents and draft proposals; a deterministic engine owns every monetary figure and every accept/reject decision, and a human gate stands before each consequential write. We realise it as six principles enforced in code, scope-limited agents writing only through validated paths, an independent second-model verifier, and a multi-week audit cycle modelled as a resumable state machine. Migrating the agents onto a model-first SDK (AWS Strands), with the engine, proposal gate, and state machine outside the model loop, shows the stance is architectural, not an artifact of a bespoke runtime. Over two quarterly cycles in production the reconciliation agreed with the regional auditors on 252 of 254 church-quarters (99.2%; 98.3% recall at 100% precision). Those auditors also use the system, so this measures agreement with current practice; controlled fault injection supplies independent evidence (100% recall on the arithmetic classes targeted, 0% on the balanced-omission class it provably cannot see), alongside a 3.21× data-entry speedup.

Paper Nr:	113
Title:	SAKE: Software Architectural Knowledge Evaluation Benchmark for Large Language Models
Authors:	Tiziano Santilli, Francesco Daghero and Mahyar Tourchi Moghaddam
Abstract:	Large Language Models (LLMs) are increasingly used as assistants across the software development lifecycle, yet their ability to reason about software architecture remains largely unmeasured. Architectural decision-making depends on quality attribute trade-offs, design patterns, and system-level constraints, none of which are exercised by benchmarks that target syntactic or algorithmic tasks. We introduce SAKE (Software Architectural Knowledge Evaluation), a standardized and reproducible benchmark for assessing software architectural knowledge in LLMs. SAKE comprises 2154 expert-curated multiple-choice questions, each with four options, stratified across eight architectural categories and four context-length levels. We evaluate 11 proprietary and open-weight models in zero-shot and five-shot settings. Overall accuracy is high, but performance varies markedly across categories, revealing competency gaps in areas central to professional practice. SAKE, its evaluation scripts, and all results are released as open source to give the community a baseline for tracking architectural reasoning in LLMs.

Paper Nr:	137
Title:	LLM-Based Multi-Agents for Sponsored-Product Campaign Management in E-Commerce
Authors:	Tiago Perrupato Antunes, Lucas Bellusci D'Andréa, Carlos Eduardo Milani Xavier, André Gomes Regino, Víctor Hochgreb de Freitas and Julio Cesar dos Reis
Abstract:	Sponsored-product campaign management requires inter- preting heterogeneous advertising signals and making budget allocation decisions that balance business objectives with marketplace constraints. Large Language Models (LLMs) offer strong reasoning for this task, but applying them to campaign management is non-trivial: recommendations must be grounded in quantitative evidence, consistent with business ob- jectives, and robust enough for operational use. We propose HMACA (Hybrid Multi-Agent Campaign Analysis), a hybrid multi-agent frame- work that pairs deterministic analytical components with specialized LLM-based agents. It decomposes campaign analysis into three cooper- ative agents, for visibility, conversion, and performance, whose outputs are integrated through controlled workflows and deterministic validation rules into product-level recommendations, separating the two layers by a single criterion: what can be checked against the data is computed deterministically, and the agents supply only the interpretation, the ac- tion, and the explanation. We evaluate HMACA on real campaigns from a large Latin American e-commerce marketplace, using a panel of in- dependent LLM judges validated against human annotations to assess recommendation quality across eight dimensions, the contribution of the deterministic guardrails, and the agreement between automated and hu- man evaluations. The results show that the pipeline performs strongest in diagnostic reasoning and evidence-to-action consistency, while business- risk awareness remains its main weakness, particularly for aggressive investment recommendations.

Paper Nr:	153
Title:	PARLEY: Agentic Wave Function Collapse through Homogeneous Agent Negotiation
Authors:	Mohd Nor Akmal Khalid
Abstract:	Agentic content generation is increasingly identified with orchestrating large language model (LLM) agents: powerful, but expensive, stochastic, and unable to guarantee that generated artifacts satisfy structural constraints. We present a contrarian alternative. Wave Function Collapse (WFC) learns local adjacency constraints from a single example, yet reasons only locally and cannot enforce the global constraints that make a game level playable: a unique player, key, and door with a path between them. We reframe generation as distributed constraint optimization solved by homogeneous, non-learning agents that negotiate; a method we call PARLEY. Every grid cell is an identical agent, and global structure emerges from negotiation at two scales: cell agents repair local incoherence by maximum-gain concession, region coalitions negotiate connective corridors, and agents bid, with losers conceding, to host the unique entities. A single temperature parameter makes the protocol genuinely stochastic, yielding a controllable quality–diversity knob at constant validity. On a Zelda dungeon domain, negotiation raises playability from at most 17% across four baselines to 100%, with a maximal effect size and about four orders of magnitude faster than search, holds 100% from 16^2 to 64^2 cells, and, as temperature rises, matches unconstrained WFC’s diversity while staying fully playable. Casting constraint solving as emergent negotiation equips a generative model with global reasoning that is fast, transparent, and verifiable.

Paper Nr:	169
Title:	BulaCheck: Verifying Claims and Questions about Medications with Package Inserts and Language Models
Authors:	Isadora de Moura Tebaldi, Jonatan de Almeida Torres Souza, Gabriel de Sapienza Luna, Rafael de Sapienza Luna, Daniel Schneider, Marcos Antonio de Almeida and Jano Moreira de Souza
Abstract:	The intense circulation of health information in digital environments has broadened public access to content about diseases, treatments, and medications, but it has also intensified the risks associated with the infodemic, including false, misleading, or decontextualized information. In the case of medications, this scenario is especially sensitive, since incorrect interpretations of indications, contraindications, adverse reactions, interactions, or ways of use can compromise the safe and adequate use of these products. Although package inserts represent a relevant source of pharmaceutical information, presenting data on composition, instructions for use, contraindications, and possible adverse effects, many users do not consult them because of the complexity of the content, the length of the documents, and the difficulty of locating specific information. Even in digital form, package inserts remain mostly static and poorly adapted to the user's informational needs. Given this problem, this work proposes BulaCheck, a system based on language models and Retrieval-Augmented Generation (RAG) to support the verification of questions and claims about medications from package inserts. From a knowledge-management perspective, the system organizes package inserts into a structured base, retrieves relevant passages, and uses this evidence to generate traceable answers, classifying claims as confirmed, refuted, or inconclusive. Thus, BulaCheck seeks to connect the conversational capacity of LLMs to reliable documentary sources, contributing to clearer, verifiable, and evidence-oriented access to information about medications.

Short Papers

Paper Nr:	15
Title:	Enhancing Kubernetes Cluster Troubleshooting Efficiency with K8sGPT
Authors:	Kalle Puutio, Mika Saari and Pekka Abrahamsson
Abstract:	Kubernetes cluster triaging is a complex and cognitively demanding process. This study examines the practical benefits and limitations of K8sGPT, an AI-powered debugging tool, through task-based interaction, semi-structured interviews, and reflexive thematic analysis. Five IT professionals, ranging from junior to senior, completed six Kubernetes debugging tasks with the assistance of K8sGPT. After one hour, participants were interviewed about their experiences, and qualitative data were analyzed to identify usage patterns. Results indicate that K8sGPT’s concise, human-readable cluster summaries reduce manual effort and cognitive load, and facilitate rapid onboarding for less experienced users. AI-generated solution suggestions enabled novice and occasional users to complete tasks they might otherwise be unable to solve. Trust in AI output increased when the tool explained its reasoning, supporting informed decision-making. However, risks included over-reliance and a tendency to address symptoms rather than root causes, particularly among less experienced users. These findings highlight both the potential and limitations of AI-assisted tools in Kubernetes operations, with possible relevance to similar complex, distributed system management tasks.

Paper Nr:	37
Title:	Agentic Incremental Method Engineering (AGIME): A Framework for Evidence-Backed Improvement of Agentic Workflows
Authors:	Ayman Asad Khan, Md Toufique Hasan, Mika Saari, Zheying Zhang, Jussi Rasku, Pekka Abrahamsson and François Christophe
Abstract:	Agentic workflows combine Large Language Model (LLM) prompts, tools, retrieval, memory, and human interaction, yet their evolution is often managed through untracked edits. Existing orchestration frameworks do not primarily treat the executed workflow specification, its runtime evidence, and its approved successor as one traceable lifecycle. Method Engineering (ME) treats methods as engineered artifacts, while Incremental Method Engineering (IME) supports their controlled evolution. Building upon that, we present Agentic Incremental Method Engineering (AGIME), a domain-neutral framework integrating a versioned Method Pack, a seven-trace Evidence Layer, and a human-in-the-loop Lifecycle Operator. AGIME executes a published, content-hash-pinned Method Pack, records typed evidence, creates a cited draft successor, and requires validation and publication before reuse. AGIME v1.0.0 operationalizes this mechanism through an executable W-kernel and passes 20 framework-derived conformance checks covering specification identity, lifecycle transitions, evidence integrity, and governed succession. In a small synthetic A-CAIRDS instantiation, the harness verifies an evidence-linked, constraint-preserving, prompt-only transition from MP- 1.0.0 to MP-1.1.0. These observations establish lifecycle feasibility under the tested fixtures, not causal improvement, operational effectiveness, or cross-domain generalization. Code and evaluation artifacts are available at https://github.com/GPT-Laboratory/AGIME-Studio under the public AGIME-1.0.0 tag.

Paper Nr:	38
Title:	Understanding Silent and Commented Rejections of Agent-Generated Pull Requests
Authors:	Mahd Hindi, Salah Bouktif and Yasir Mahmood
Abstract:	Agentic coding tools are increasingly able to modify repositories, generate patches, and submit pull requests (PRs) to real software projects. However, many agent-generated pull requests (APRs) are rejected by maintainers or repository-level gates, raising an important question: what can these rejections reveal about the limitations of current coding agents? This paper presents a focused empirical study of rejections in APRs. Using the AIDev-POP subset of the AIDev dataset, we analyze 12,433 APRs submitted by five coding agents across 1,495 popular GitHub repositories. We distinguish between silent rejections, where rejected APRs are closed without inline reviewer feedback, and commented rejections, where reviewer comments provide visible evidence of failure. We find that 84.2% of rejected APRs are silent, revealing a substantial visibility gap within the analyzed rejected-APR (RAPR) cohort and limiting the observable evidence available for explaining these outcomes. For the commented rejections, we apply a validated taxonomy-based classification pipeline to 3,750 review comments associated with 676 rejected APRs. The results show that functional failures, particularly specification mismatches and logic defects, are the dominant visible rejection pattern. These findings suggest that future coding agents should become more reflexive by acquiring repository context, validating changes before submission, and providing reviewer-grade evidence to support human review.

Paper Nr:	62
Title:	ARTEMIX Verification-First: A Closed-Loop Evidence-Validity Framework for Agentic Electronic Hardware Engineering
Authors:	Saeid Jamili, Sabereh Taghdisi Rastkar, Stefano Perticaroli and Mauro Olivieri
Abstract:	Electronic hardware engineering relies on artifacts such as schematics, layouts, RTL blocks, firmware, simulation reports, timing results, board measurements, hardware-in-the-loop (HIL) traces, and outputs produced by agentic AI systems. These artifacts support an engineering claim only when the evidence behind them is valid for the configuration under review. In practice, evidence that once passed may become stale after changes in design files, extracted views, package models, PDKs, tools, constraints, firmware, bitstreams, workloads, hardware setups, HIL fixtures, review decisions, accepted-risk records, or tool-interface policies. This paper proposes ARTEMIX Verification-First, a closed-loop evidence-validity framework for agentic electronic hardware engineering. Its central object is the Evidence Validity Contract (EVC), which links an engineering claim to an artifact, verification method, result, validity boundary, invalidation triggers, review state, decision state, accepted risk, and learning feedback. The framework defines evidence debt, evidence gates, engineer-in-the-loop review, hardware-coupled validation, responsibility separation, and learning-to-planning feedback. HIL testing, FPGA-in-the-loop, emulation, and prototype execution are treated as bounded evidence sources for claims that depend on timing, I/O behavior, physical interaction, or system-level effects. Model Context Protocol (MCP)-compatible interfaces are considered as one way to connect agents with tools and resources, but tool invocation is not release permission, and release authority remains with responsible engineers. ARTEMIX AI is used as an implementation architecture that illustrates task contracts, evidence packets, claim ledgers, engine

Paper Nr:	116
Title:	Controlled Agentic Patching for Production Workflow Graphs
Authors:	Aleksandra Mikheeva
Abstract:	Large language model (LLM) coding agents resolve many repositorylevel issues, yet production software maintenance exposes a failure surface that isolated code generation does not capture. This position paper addresses agentic debugging and patching of production workflow graphs: directed graphs in which data, machine-learning, forecasting, and validation pipelines are expressed as interdependent tasks with generated code, configuration, schemas, runtime parameters, logs, and historical run metadata. We argue that final patch correctness is an insufficient acceptance criterion; production readiness additionally requires reproducible context, validation coverage, rollback safety, blast-radius control, and review quality. Rather than a conceptual warning, the paper is an operational specification for production-ready agentic patching, delivered as concrete artifacts: a failure-mode taxonomy, a Controlled Agentic Patching Loop, an explicit autonomy boundary with a per-capability permission policy, deterministic validation gates, an evidence-report schema, a retrospective incident-analysis protocol with production-readiness evaluation criteria, a representative industrial incident pattern, and a catalogue of implementation frictions. The agent inspects context, localizes a root cause, proposes a bounded patch, runs validation gates, emits a reviewer-facing evidence report, and requests human approval; it never merges or deploys. We separate patch generation from patch production readiness through gates with explicit escalation conditions, and we define evaluation criteria that score localization, validation completeness, escalation correctness, reproducibility, and blast-radius violations rather than test pass rate alone.

Paper Nr:	139
Title:	The Agentic Company OS: Substrate Inversion for Sustained Enterprise Agent Deployment
Authors:	Oliver Aleksander Larsen and Mahyar T. Moghaddam
Abstract:	Enterprise AI agents often succeed in a demonstration and then stall once they must operate day after day. An industry report estimates that most pilots never reach production and that deployed systems rarely retain feedback or improve over time, while agent benchmarks show single-run successes masking unreliable repetition. We argue that these failure modes share a common architectural root: agents reason over data structured for human operators and traditional applications, not for the language models that power them. This position paper proposes that companies deploying agents in sustained operation should rebuild their cognitive substrate, the shared environment agents read as working context, around representations matched to that reasoning surface, isolating schema translation to the action boundary. Markdown is the instantiation available today, not a proven agent-native primitive. Two mechanisms ground the argument: context-bandwidth asymmetry, the gap between one-pass reading of connected prose and field-by-field typed access that strips relations; and cross-loop coupling, the claim that action, skill, and policy loops compound only if they share one substrate. A four-layer framework (Data, Knowledge, Intelligence, Governance) operationalizes the position, with a Sync Agent enforcing the action boundary and a per-skill trust gradient, making governance and auditability structural properties of the substrate. The position revives the shared-substrate tradition of classical multi-agent systems under LLM-era economics. We analyze the main objections and risks, including indirect prompt injection on the compile path, and outline a research agenda for evaluating substrates directly.

Paper Nr:	165
Title:	InnoSpire Glasses Companion: An AI-Driven Multilingual Assistive Mobile Application for Visually Impaired and Blind Individuals
Authors:	Zack Gluck, Samruddhi Naik, Chun-Kit Ngan, Kai-Wing Siu and Ricky Chan
Abstract:	In this work, we propose the design and development of a multilingual assistive mobile application integrated with InnoSpire Glasses to support visually impaired and blind (VIB) individuals in supermarkets, restaurants, and public transportation settings. The key contributions of this work are five-fold: (1) a voice processing module for hands-free interaction that includes (i) keyword spotting, (ii) voice activity detection, and (iii) automatic speech recognition; (2) a seamless switching mechanism that automatically transitions between online mode using Google Gemini 2.5 Flash and offline mode using SmolVLM2 based on network connectivity; (3) automatic language translation among English, Hindi, Mandarin, and Cantonese; (4) an accessible user interface designed in accordance with the Web Content Accessibility Guidelines and the Americans with Disabilities Act standards; and (5) a comprehensive evaluation of our mobile application through API benchmarking, comparison with existing VIB applications, and assessment by VIB participants. In our API evaluations, Google Gemini 2.5 Flash achieved the second-highest LLMScore of 0.93, compared to 0.91 for Seeing AI and 0.87 for Be My Eyes, while maintaining one of the lowest costs at \$0.372 per 1,000 requests and a fast average response time of 4.9 seconds. In the user study, the overall ease of use of our application was rated 3.9 out of 5, and the overall quality of app responses received an average rating of 4.2 on a 5-point Likert scale.

Paper Nr:	170
Title:	An Agentic Pipeline of Language-Model Agents, Tools, and Knowledge Graphs for Evidence-Grounded Verification of Medical Claims
Authors:	Jonatan de Almeida Torres Souza, Isadora de Moura Tebaldi, Gabriel de Sapienza Luna, Rafael de Sapienza Luna, Daniel Schneider, Marcos Antonio de Almeida and Jano Moreira de Souza
Abstract:	The biomedical literature grows at a pace that makes the manual verification of clinical claims infeasible, while large language models (LLMs), though fluent, are prone to hallucination and offer little traceability—a critical shortcoming in high-stakes domains such as health. We investigate how structured knowledge, in the form of knowledge graphs, can be combined with LLMs to verify medical claims with traceable evidence, and we frame the resulting system as an agentic pipeline: a set of reasoning-and-acting language-model agents coordinated with deterministic tools that query authoritative biomedical sources. As a case study, we present a seven-stage pipeline that, from a claim written in Portuguese, retrieves PubMed articles live through MeSH terms, builds per-claim and per-article knowledge graphs, and produces an explainable verdict by confronting them. The approach was evaluated on a balanced set of 40 claims and compared against a pureLLM baseline across four distinct language models (two open-weight, run locally, and two proprietary, accessed via API), using the F1-score as the primary metric. The results show that the baseline attains a high F1, but its near-perfect agreement with an automatically generated gold standard reflects shared, non-independent bias among language models rather than demonstrated factual correctness; the proposed pipeline instead abstains when the retrieved evidence is inconclusive, which lowers its F1 under a strict scoring rule yet yields, by design, intermediate graphs that make each verdict traceable and auditable. We complement the comparison with an abstention-aware analysis, an error analysis, worked examples, and a discussion of the threats to validity, and we point to directions for refining the generation and comparison of the graphs.

Paper Nr:	40
Title:	TraceCoder: Explainable and Auditable Code Generation
Authors:	Rwaida Alssadi, Muntaser Syed, Balaji Kasula, Lamine Deen, Majed Alotaibi, Mohammed Alghamdi, Tyler Ton, Ali Alqarni and Marius Silaghi
Abstract:	Contemporary LLM-based coding agents produce code as black-box outputs: the faithful rationale behind each line is hidden except as stochastically generated in associated comments, the evolution of the code through benchmarkdriven repair is ephemeral, and post-hoc auditing is impossible. We present a code generation concept based on benchmark tests and improvement iterations that addresses these shortcomings through three complementary mechanisms: (i) a relational snippet-history schema that records, per repair event, the benchmark reference, round number, failure text, and LLM explanation, enabling full provenance queries; (ii) a browser-based visualisation tool that renders this history as heat-mapped, hover-annotated source code; and (iii) a competitive fractional position-key indexing scheme with tree-node delimiters that assigns stable, lexicographically-ordered identifiers to each code snippet, enabling fine-grained tracking without disrupting surrounding lines. While the auditability claim is theoretically justified, we empirically evaluated TraceCoder’s usability on 30 simple algorithmic programming tasks spanning string processing, mathematical computation, and data-structure manipulation, across two provider configurations. Of these, 10 exhaust the 6-iteration budget on tasks with subtle edge-case behaviour. Mean fraction of snippets with at least one benchmark failure explanation reaches 30% with different LLM providers for code and benchmarks, compared to 21% when using a sole provider on a 20-task subset. Three detailed case studies demonstrate how the system explains which specific benchmark failures shaped each line of the final program. The proposed mechanism makes the internal “narrative” of automated code generation auditable and replayable, a property essential for trust and accountability in production deployments.

Paper Nr:	63
Title:	Agents4experiments: LLM Agents for Physical Experiment Design
Authors:	Chan-Eui Song, Shohom Bose-Bandyopadhyay, Gyunghun Yu, Peter Pak, Francis Ogoke and Amir Barati Farimani
Abstract:	Large language models (LLMs) have demonstrated extensive capability in code generation and simulation-based research workflows, however, their ability to design physically feasible experiments remains underexplored. This is of particular significance in engineering related fields such as additive manufacturing, welding, and battery design, where progress depends on executable physical experiments. In this work, a bill-of-materials (BOM)-grounded agentic system is developed and evaluated against a standard GPT-5 baseline. A literaturegrounded database is constructed by converting retrieved paper chunks into structured science cards and experiment cards, and used to construct a multi-agent retrieval-augmented generation system, an elsapy-driven retrieval pipeline, and fine-tuned models based on proposal trajectories. All approaches are evaluated using a validator agent grounded in a Design of Experiments rubric. Results show that larger models produced stronger proposals overall, while criterion-level analysis revealed persistent weaknesses in structured DoE reasoning. Query-specific retrieval improved clarity and compliance, but retrieval and fine-tuning did not eliminate these attribute-level failure modes. These findings suggest that structured DoE reasoning remains a key bottleneck for LLM-based physical experiment design.

Paper Nr:	167
Title:	HEAL: Human-Expert Assisted Learning for Collaborative Medical Agents
Authors:	Rafael Luna, Gabriel Sapienza Luna, Daniel Schneider, Marcos Antonio de Almeida, Carlos Eduardo Barbosa, Matheus Argôlo and Jano Moreira de Souza
Abstract:	Healthcare workforce strain is a global challenge associated with long waiting times, limited consultation duration, and increasing burnout among healthcare professionals. This paper presents HumanExpert Assisted Learning (HEAL), a multi-agent framework for investigating evidence-grounded clinical interactions and supporting the future integration of expert feedback. HEAL combines a Patient Persona Generator, a Patient Simulator, a Clinician Agent, and an Evaluation Agent with a Retrieval-Augmented Generation (RAG) pipeline grounded in the Clinical Protocols and Therapeutic Guidelines of the Brazilian Unified Health System. The framework was evaluated using 50 LLM-generated synthetic clinical cases and ten language-model configurations under a standardized dialogue protocol with a maximum of ten anamnesis questions. Model outputs were assessed by an LLM-based evaluator according to scientific-consensus alignment, clinical reasoning, automated safety scores, and potentially severe warning-sign omissions identified by the evaluator. The highest observed mean scores were obtained by gpt-oss:20b and gemma3:12b, while qwen2.5:7b and deepseek-r1:8b also showed comparatively strong performance despite their smaller parameter counts. The observed ranking was not monotonic with model size. The current study constitutes a preliminary automated benchmark rather than clinical validation: the synthetic reference diagnoses were not independently verified by physicians, and the Evaluation Agent was not calibrated against expert medical judgments.