MAS-GAIN 2026 – Preliminary Table of Contents

2nd International Workshop on Multi-Agent Systems using Generative Artificial INtelligence for Automated Software Engineering (MAS-GAIN 2026)

Frontmatter

Papers

Does Architecture Matter? Comparing Multi-agent Workflows for Filtering and Prioritizing Static Analysis Findings
Linus Bennin, Oshando Johnson, and Eric Bodden
(TU Darmstadt, Germany; Fraunhofer IEM, Germany; Paderborn University, Germany)
The number of exploited software vulnerabilities rises every year, and their impact grows as software systems become ever more interconnected. To detect such vulnerabilities in large codebases, many organizations rely on Static Application Security Testing (SAST) tools. However, SAST tools can produce high volumes of false positives, which increases manual triage effort, causes alert fatigue, and can lead developers to dismiss warnings that contain real vulnerabilities. Large Language Models (LLMs) and AI agents are increasingly used to filter such findings, and it is already well established that newer, stronger models yield better results. Less is known about an orthogonal design choice: does the multi-agent workflow architecture itself change filtering performance? We implement a modular multi-agent framework comprising six agents and four distinct workflow architectures: single-agent, sequential pipeline, parallel debate, and deep agent. We evaluate the architectures on two synthetic open-source benchmarks and one real-world benchmark containing actual Common Vulnerabilities and Exposures (CVEs) in full repositories, using three different models. The evaluation results show that, for a fixed model, the choice of architecture notably shifts filtering performance, with the Multi-Agent Debate architecture achieving the strongest filtering, albeit also at the highest token cost. Moreover, the workflows' design decisions, such as passing prior reasoning between agents or looping findings back for re-auditing, govern a controllable trade-off between false-positive reduction and recall. We further show that tool-aware retrieval improves recall without harming precision, and that agent-generated context makes vulnerability prioritization markedly more stable and steerable.
Article Search Article: asews26masgainmain-id7-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844516

Decompose to Conquer: A Multi-agent LLM Pipeline for Requirements Engineering and Software Design
Mesfin Tibebu and Mohamed Nabeel
(National University, USA)
The automation of Requirements Engineering (RE) and software design using Large Language Models (LLMs) remains an open challenge, particularly in capturing stakeholders’ intent accurately and fully for complex real-world software. In this work, we investigate whether decomposing the RE and software design phases of the Software Development Life Cycle (SDLC) into specialized collaborative agents, rather than relying on a single monolithic LLM call, yields measurably better design quality, and whether grounding those agents in a domain-specific Retrieval-Augmented Generation (RAG) knowledge base provides significant additional gains over general-purpose retrieval. To this end, we propose READ-MAS, an LLM-based multi-agent system (MAS) comprising five specialized agents backed by a RAG pipeline, and evaluate it through a quantitative constructive study. First, READ-MAS’s multi-agent architecture, comprising five specialized agents for RE (Collector, Analyzer, Specifier) and design (Designer, Documenter), each handling a discrete sub-task, statistically significantly outperforms a monolithic single-agent baseline (one LLM call handling the full RE and design task in a single pass) on design accuracy (Welch’s ANOVA, 𝑝 = 0.018, Hedge’s 𝑔 = 1.19) and RAGAS faithfulness (Kruskal-Wallis, 𝑝 < 0.001, multi-agent median = 0.999 vs. single-agent median = 0.851), while achieving equivalent downstream coding performance on HumanEval and MBPP. Second, equipping READ-MAS with a domain-specific RAG knowledge base yields a statistically significant improvement in design accuracy over RAG-disabled runs (One-way ANOVA, 𝑝 = 0.002, mean difference = 2.65%), while a general-purpose knowledge base fails to produce a significant gain (𝑝 = 0.266). These results provide empirical evidence that (a) the decomposition of RE and the design into specialized agents is likely to increase the quality of the design beyond what a single agent can achieve, and (b) the effectiveness of RAG is contingent on the specificity and domain relevance of the knowledge-base rather than mere retrieval augmentation
Article Search Article: asews26masgainmain-id16-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844517

Towards Declarative Composition of Heterogeneous AI Agents
Antonio Marcos Alves Morais, Paulo Henrique Mendes Maia, and Felip Manyà
(State University of Ceará, Brazil; IIIA-CSIC, Spain)
As AI systems increasingly evolve into ecosystems of specialized agents, composing heterogeneous AI technologies into coherent multi-agent applications has become a significant software engineering challenge. Existing agent composition frameworks primarily focus on large language models and conversational workflows, leaving classification, regression, and other AI services tightly coupled to framework- or provider-specific implementations. This limits portability, interoperability, and the consistent governance of heterogeneous agent ecosystems. We propose a declarative composition model in which AI agents are specified as portable, contract-governed artifacts independent of the underlying AI technology, model, or provider. The model separates agent declaration, dispatch, and execution, shifting integration concerns from imperative orchestration code to declarative specifications while enabling uniform runtime validation, governance, and provider-agnostic dispatch. We materialize the proposed model in a runtime architecture supporting agent registration, contract enforcement, Agent Level Agreements (ALAs), telemetry, and transparent model substitution. A reproducible feasibility study with real LLM-based code-smell analysis agents demonstrates that heterogeneous agent roles can share the same composition pipeline while preserving contracts, runtime policies, and portability across model configurations. Although evaluated using generative agents, the proposed model is designed as a software engineering foundation for composing and governing heterogeneous AI agent ecosystems.
Article Search Article: asews26masgainmain-id17-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844518

Towards Reliable Agentic Infrastructure Provisioning with LLMs
Yhon Castañeda and Camilo Escobar-Velásquez
(Universidad de los Andes, Colombia)
Kubernetes (K8s) is widely used to orchestrate containerized workloads, yet provisioning reliable and maintainable clusters requires substantial DevOps and cloud-infrastructure expertise. This paper presents an LLM-based Infrastructure-as-Code workflow for Amazon Elastic Kubernetes Service (EKS) that combines retrieval-augmented generation (RAG) with a bounded validation-and-repair loop. The workflow retrieves contextual examples from a curated Terraform knowledge base, generates infrastructure configurations, validates them, and returns recoverable errors to the model for correction. We evaluate the approach through two complementary studies. First, an exploratory infrastructure-provisioning benchmark assesses six commercial LLMs across five prompt variants using the complete retrieval-and-repair workflow to provision AWS EKS infrastructure. Second, a controlled 2 × 2 ablation independently enables and disables retrieval and iterative repair for three representative models. This ablation includes five prompt variants and three repetitions per model-condition combination, totaling 180 plan-only executions. Under the complete workflow, Terraform plan success reached 100% for Claude Opus 4.5, 93.3% for GPT-4.1, and 100% for Gemini 3.1 Pro, compared with one-shot baseline rates of 86.7%, 0%, and 46.7%, respectively. The results show that retrieval and repair have model-dependent effects, while their combination achieved the highest or joint-highest plan success for all reevaluated models.
Article Search Artifacts Available Article: asews26masgainmain-id12-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844519

IaC agentic workflow (doi:10.5281/zenodo.22652454): The artifact contains the source code, experimental configurations, prompt variants, evaluation scripts, and supporting evidence for an LLM-based Terraform workflow that combines retrieval-augmented generation and iterative validation and repair to provision Amazon EKS infrastructure. It supports the exploratory ...

Neuro-symbolic Orchestration for Autonomous Travel Execution: Mitigating LLM Hallucinations via Deterministic Constraints
João Cordeiro and Piedade Carvalho
(Institute of Engineering of Porto, Portugal)

Article Search Article: asews26masgainmain-id2-p (type: Short Paper (4 pages)) doi:10.1145/3843778.3844520

SAGE: Source-Aligned Governance for Human–Agent Requirements Clarification in Multi-agent Software Engineering
Tingting Feng, Meifang Xue, Jian Wang, and Zhongyan Du
(China Unicom Smart City Research Institute, Beijing, China; Zhonggongjinglian Talent Service, Zhejiang, China)
Multi-agent large language model workflows are increasingly used in software engineering, yet requirements clarification under human–agent collaboration remains weakly governed. Much of the rework reflects requirement-chain desynchronization: negotiated intent, workflow contracts, and signed deliverables evolve on separate tracks without auditable linkage, yielding semantic drift and unclear accountability. Source-Aligned Governance and Evolution governs clarification via three co-evolving sources, a clarification–confirmation–sign-off path, three strategies for cross-source mapping, staged human sign-off, and cross-scenario reuse, and ten operational mechanisms—platform-agnostic rules that encode them as auditable transition constraints—implemented as declarative transition guards for multi-agent coordination. On a reference platform, the Product Requirement to Initial Demo and Build Scene scenarios supply a built-in thermal-power plant smart-overhaul sample for artifact-level verification of sign-off and rollback on requirements artifacts, plus in-organization trial reception. The online prototype extends prior agent pipelines with executable upstream requirements guards on bound requirement artifacts. The contribution is encodability and traceability of governance constraints, not causal effectiveness.
Article Search Article: asews26masgainmain-id6-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844521

SAGE prototype operator guide and PRID workflow contracts: Supplementary material for the SAGE demonstration. The package includes the online-prototype operator guide used at review (Appendix B) and the twin PRID workflow contracts (workflow.json and workflow-task-desc.json) that bind phases, evaluation gates, and docId identifiers as discussed in the paper. An English-UI ...

Sufficiently Advanced Agents Are Indistinguishable from Erlang: An Operational Mapping from OTP to Multi-agent LLM Systems
Mick Darling
(Dollhouse Research, USA)
Multi-agent LLM systems are rediscovering the engineering problem that produced Erlang/OTP: how to build systems more reliable than their individually unreliable parts. We treat this convergence as a predictive engineering model rather than an analogy. We develop a term-by-term correspondence between OTP constructs and agent systems, grading twenty mappings as exact, loose, or breaking; the breaking mappings mark the real limits of the transfer: context isolation, payload trust, and mid-inference preemption. The correspondence draws on production experience: Erlang ingest fleets processing millions of tweets per hour, and our present agent runtime's independent reproduction of failure shapes OTP was built to survive: orphaned workers, token-burning polling loops, and sessions held open across hour-scale delays. We then test supervision as a capability multiplier in preregistered experiments that cross supervision architecture with model tier. Across 2,400 episodes and eight model tiers, OTP-disciplined supervision never loses to bare agent loops, leads at four of the five tiers where the arms separate, and shows its largest edge where models partially succeed (pass^4 at 3B: 95% supervised vs. 65% bare; at 32B: 70% vs. 45% with retry and 5% without). A preregistered matched-compute control and a prospective replication (640 episodes in total) rule against the evaluated best-of-N control as the explanation: supervision beat nominally token-matched best-of-N in both (paired deltas +0.33 and +0.26, p <= 0.004), and the preregistered selector added zero successes over its first candidate; compute parity itself was not certified in either run. On hard decomposable tasks, supervised clean-context decomposition lifts a 32B model from 57% to 82%, monotonically from one to sixteen LLM actors, on a prospectively frozen holdout. We close with the problems our work to date leaves open: semantic failure detection, context leakage, and restart economics.
Article Search Article: asews26masgainmain-id10-p (type: Full Paper (8 pages)) doi:10.1145/3843778.3844522

Mapping Evidence and Aggregate Results: Auxiliary material for “Sufficiently Advanced Agents Are Indistinguishable from Erlang: An Operational Mapping from OTP to Multi-agent LLM Systems.” It includes a three-page mapping-evidence supplement; row-level evidence and framework-by-construct matrices in CSV and XLSX formats; aggregate results underlying Tables ...

A Workflow-Driven Multi-agent Architecture for Requirements Engineering
Radosław Klimek
(AGH University of Krakow, Poland)
Requirements Engineering (RE) is an iterative process involving interpretation, clarification, modelling, and validation. Existing LLM-based approaches often rely on conversational interaction or specialised agents without explicit control over the progression and exchange of intermediate artefacts. We present a workflow-driven multi-agent architecture that coordinates specialised RE capabilities through explicit workflow states and structured artefacts, with human clarification and validation as explicit control points. An exploratory evaluation against a conversational baseline examines clarification and correction effort, while ablation analysis assesses the contribution of explicit clarification and validation. The results indicate fewer interaction cycles for the evaluated scenarios and show that removing these control points impairs convergence. The architecture provides a transparent and controllable basis for progressing from informal requirements towards behavioural models and downstream formalisation.
Article Search Article: asews26masgainmain-id9-p (type: Short Paper (4 pages)) doi:10.1145/3843778.3844524

proc time: 0.22