Artificial intelligence has moved beyond models and inference into the realm of AI Engineering, the discipline of building and operating complex AI systems at scale. From autonomous coding agents to context‑aware developer tools, the last few months saw an explosion of research that will shape how software is developed, tested and maintained. The AI Engineering research report, September '25 edition - unpacks the most significant papers, organized by theme, and highlights what each contribution means for practitioners who want to stay on the cutting edge of AI‑powered engineering.

Papers Covered in This Article

1. Agentic AI for Software: Thoughts from the Software Engineering Community
2. On the Future of Software Reuse in the Era of AI Native Software Engineering
3. SWE‑Effi: Re‑Evaluating Software AI Agent System Effectiveness Under Resource Constraints
4. LightAgent: Production‑Level Open‑Source Agentic AI Framework
5. app.build: A Production Framework for Scaling Agentic Prompt‑to‑App Generation
6. MultiFluxAI: Enhancing Platform Engineering with Advanced Agent‑Orchestrated Retrieval Systems
7. VisDocSketcher: Towards Scalable Visual Documentation with Agentic Systems
8. Context Engineering for Multi‑Agent LLM Code Assistants Using Elicit, NotebookLM, ChatGPT and Claude Code
9. AI‑Guided Exploration of Large‑Scale Codebases
10. Breaking Barriers in Software Testing: The Power of AI‑Driven Automation
11. When the Code Autopilot Breaks: Why LLMs Falter in Embedded Machine Learning
12. An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems
13. Designing an Interdisciplinary Artificial Intelligence Curriculum for Engineering
14. Teaching at Scale: Leveraging AI to Evaluate and Elevate Engineering Education
15. Agentic AI for Software: Thoughts from the Software Engineering Community

Agentic Software Engineering: From Vision to Frameworks

A roadmap for agentic software engineering

AI agents are poised to become full‑fledged members of software teams. In Agentic Software Engineering: Foundational Pillars and a Research Roadmap, Ahmed Hassan and colleagues articulate a vision for SE 3.0, where intelligent agents handle complex, goal‑oriented software engineering tasks rather than simple code completion. They propose two complementary workbenches: an Agent Command Environment, where humans orchestrate and mentor agent teams, and an Agent Execution Environment for agents to perform tasks while calling on humans when necessary. The paper lays out a research agenda that emphasises structured human‑agent collaboration, new processes such as merge‑readiness packs and consultation requests, and a vocabulary to catalyse community discussion

Thoughts from the software engineering community

In Agentic AI for Software: Thoughts from the Software Engineering Community, Abhik Roychoudhury argues that agentic AI should encompass much more than prompt‑based code generation. He sketches a future where AI agents assist with requirements understanding, architecture exploration, testing and program repair. The key challenge, he notes, is intent inference—deciphering the developer’s goals so that agents can make autonomous micro‑decisions without losing trust. The article calls for research into AI‑based verification and validation and highlights the need for agentic workflows that integrate program analysis tools.

Reusing software in the AI native era

Software reuse is changing as developers rely on AI‑generated code. On the Future of Software Reuse in the Era of AI Native Software Engineering warns that trusting generative AI can lead to a form of “cargo‑cult” development. Antero Taivalsaari and colleagues discuss the implications of AI‑assisted reuse and outline a research agenda for tackling its central issues. They argue that while reuse promises productivity gains, it also raises questions about code provenance, quality and maintainability.

Evaluating AI agents under resource constraints

Accuracy alone isn’t enough—AI agents must also be efficient. SWE‑Effi: Re‑Evaluating Software AI Agent System Effectiveness Under Resource Constraints introduces new metrics that balance solution accuracy with resource consumption. By re‑ranking systems on the SWE‑bench benchmark, the authors show that “token snowball” effects and “expensive failures” can consume excessive compute without improving outcomes. Their analysis reveals trade‑offs between token budget and time budget and highlights the need for cost‑aware evaluation.

Light‑weight frameworks for multi‑agent systems

Building agentic systems requires tooling. LightAgent: Production‑Level Open‑Source Agentic AI Framework presents a minimal yet powerful platform that balances flexibility and simplicity. The framework combines memory modules, tool generators and a Tree‑of‑Thought reasoning engine within a ~1 KLOC codebase. It supports autonomous learning, error detection and multi‑modal data handling, and offers automated generation of hundreds of domain‑specific tools. LightAgent aims to democratise multi‑agent systems by making them easy to deploy across research and industry.

Production frameworks and retrieval systems

Several papers propose concrete platforms for deploying agentic AI. app.build: A Production Framework for Scaling Agentic Prompt‑to‑App Generation with Environment Scaffolding describes an open‑source stack that combines multi‑layer validation, stack‑specific orchestration and model‑agnostic architecture. By introducing structured environments, the framework achieves a 73.3 % viability rate across 30 application‑generation tasks and demonstrates that open‑weights models reach 80.8 % of closed‑model performance when provided proper scaffolding. MultiFluxAI: Enhancing Platform Engineering with Advanced Agent‑Orchestrated Retrieval Systems introduces an AI platform that leverages generative AI, vectorisation and agentic orchestration to integrate disparate data sources and provide context‑aware responses. While the abstract is brief, it signals a trend toward using agents to manage complex retrieval and engineering workloads.

Automatic visual documentation

Code visualisation helps developers understand unfamiliar systems, but creating diagrams is labour‑intensive. VisDocSketcher: Towards Scalable Visual Documentation with Agentic Systems offers the first agent‑based method for automatically generating visual documentation. The system combines static code analysis with LLM agents to identify key elements and produce visual sketches. The authors also propose AutoSketchEval, a framework that uses code‑level metrics to assess visual documentation quality, achieving an AUC above 0.87. This work lays the groundwork for automated visual documentation and suggests that agentic systems can reduce cognitive load for developers.

AI Agents in the Developer’s Toolkit

Context engineering for multi‑agent code assistants

Large‑language‑model code assistants struggle with complex, multi‑file projects. Context Engineering for Multi‑Agent LLM Code Assistants Using Elicit, NotebookLM, ChatGPT and Claude Code introduces a workflow that clarifies user intent, retrieves domain knowledge, synthesises documents and orchestrates specialised sub‑agents The multi‑agent system improves code generation accuracy by injecting context via semantic retrieval and document synthesis, outperforming single‑agent baselines and highlighting the importance of context management in LLM‑powered assistants.

Interactive exploration of large codebases

Understanding complex software is a perennial challenge. AI‑Guided Exploration of Large‑Scale Codebases proposes a hybrid tool that combines deterministic reverse engineering with LLM‑guided, intent‑aware visual exploration. Developers can use UML‑based visualisations and dynamic user interfaces to navigate code, while the LLM interprets queries and interaction patterns to suggest relevant parts of the system. A prototype for Java demonstrates the feasibility of integrating structured views with LLM guidance.

AI‑driven software testing

Testing remains a bottleneck in software development. Breaking Barriers in Software Testing: The Power of AI‑Driven Automation introduces a framework that translates natural language requirements into executable test cases using NLP and reinforcement learning, optimises them through continuous learning, and validates results with real‑time analysis. By embedding these techniques within a trust and fairness model, the authors report improved defect detection, reduced testing effort and faster release cycles. This work illustrates how AI can shift testing from manual to proactive, adaptive processes.

When the code autopilot breaks

Large‑language‑model pipelines can silently fail. When the Code Autopilot Breaks: Why LLMs Falter in Embedded Machine Learning investigates failure modes in LLM‑powered embedded ML workflows. The authors analyse an “autopilot” framework that orchestrates data preprocessing, model conversion and on‑device inference. They show that prompt format, model behaviour and structural assumptions influence success rates and expose error patterns that standard validation does not catch. The paper derives a taxonomy of failure categories and urges more robust validation and traceability.

Parameter estimation via agentic workflows

Beyond software, AI agents can simplify scientific workflows. An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems presents a pipeline that converts a human‑readable problem description into a compiled, differentiable calibration pipeline using JAX and automatic differentiation. The system automatically validates consistency between specification and code, auto‑remediates pathologies and orchestrates a two‑stage search with global and gradient‑based optimization. By lowering the barrier to calibrating mechanistic ODE models, this agentic workflow demonstrates how AI engineering can accelerate scientific discovery.

AI Engineering in Education and Curriculum

Designing an AI curriculum for engineers

As AI reshapes engineering practice, education must adapt. Designing an Interdisciplinary Artificial Intelligence Curriculum for Engineering examines a novel undergraduate program that integrates AI competencies across disciplines. Using curriculum mapping and focus‑group interviews, the authors assess alignment with targeted skills and evaluate perceived quality, practicality and effectiveness from academic and industry perspectives. The study highlights the importance of educator participation in curriculum development and offers insights for universities designing AI‑native engineering programs.

Evaluating teaching at scale with AI

Large engineering programs struggle to synthesise qualitative student feedback. Teaching at Scale: Leveraging AI to Evaluate and Elevate Engineering Education proposes an AI‑supported framework that uses hierarchical summarisation and anonymisation to extract themes from open‑ended comments. Visual analytics contextualise numeric scores through percentile comparisons and historical trends, while ethical safeguards ensure privacy The system has been deployed across a large engineering college, and preliminary validation suggests that LLM‑generated summaries can reliably support formative evaluation.

Cross‑Cutting Themes and Emerging Trends

Several patterns emerge from these papers:

• Human–agent collaboration: Research emphasises collaborative workflows where agents assist rather than replace engineers. Frameworks like SE 3.0, LightAgent and app.build design shared environments and tools that enable bi‑directional interaction.
  
  
• Context and retrieval: Multi‑agent code assistants and platform engineering systems highlight the importance of context engineering and retrieval‑augmented generation. Properly injecting domain knowledge dramatically improves agent performance.
  
  
• Efficiency and resource awareness: Papers such as SWE‑Effi and When the Code Autopilot Breaks show that evaluating AI systems requires balancing accuracy with cost and identifying failure modes.
  
  
• Automation across the software lifecycle: From generating visual documentation to automating parameter estimation, AI agents are poised to streamline tasks across development, testing and maintenance.

Explore More on AI Engineering

If you’re interested in diving deeper into AI Engineering and related topics, here are some of our recent articles on AryaXAI that complement this research review:

• The AI Interpretability Research Review, September ’25 Edition: A Foundational Leap in Model Interpretability - Covers analysis of latest model interpretabillity research papers what model interpretability means, proposes new architectures with built‑in transparency, and tackle domain‑specific problems ranging from time series forecasting to credit scoring.
• From Abstract Theory to High-Stakes Application : The Alignment Report (September '25), analysis of latest AI Alignment research papers and their findings.‍
• Latest AI Research Papers: July 2025 Roundup — Part 2, which analyses cloud‑native inference stacks, heterogeneous orchestration and hardware co‑design strategies that underpin modern AI systems.‍
• AI Alignment vs. Model Performance – How to Optimize for Accuracy, Compliance, and Business Goals, discussing how to balance predictive power with safety and governance.

These resources complement the latest research and provides additional context for building AI systems that are not only powerful but also trustworthy, efficient and transparent.