We summarize findings from analyzing developers’ responses to a prompt regarding how AI-powered tools have impacted their recent development workflow. We identify four tensions that developers must balance in their use of AI and suggest strategies practitioners can use to improve developer experience with AI.
This article explores whether learning to code is still necessary in the age of AI. Drawing on survey responses, teaching experience, and personal reflection, it examines how AI coding assistants are changing software development, the skills future technologists will need, and the importance of maintaining technical understanding in an increasingly automated world.
This article explores the human-centric factors shaping software testers’ adoption of large language model agents. Drawing on a survey, we propose an adoption framework comprising six factors to help testers evaluate and redesign testing practices for the long-term adoption of LLM agents.
In this article, a new approach based on gated hierarchical bidirectional long short-term networks is proposed for software defect prediction. The experimental results show that the proposed approach outperforms existing methods in a within-item scenario.
This “Practitioner’s Digest” column highlights emerging software engineering for AI research presented at the IEEE/ACM International Conference on Software Engineering 2026, focusing on reliability, runtime safety, sustainability, requirements alignment, and scalable AI operations.
Modernizing legacy software increasingly depends on code translation tools that migrate applications across languages, frameworks, and platforms. Here we compare the current generation of AI-based and rule-based transpilers, analyzing their supported ecosystems, integration models, and engineering tradeoffs to help practitioners understand when lightweight large language model-based translation is sufficient and when deterministic modernization platforms remain necessary.
A Git extension identifies source-code hotspots by analyzing code churn and line lifetimes across repository histories. Its development illustrates extending Git, transforming a research prototype into a Python package, integrating shell pipelines, configurable output formatting, and lessons from agentic AI software development.
For this first guest editorial, we get an update from a Requirements Engineering Workshop discussing how AI will impact the future of their field.
In this column, the author takes us through the integration of disparate poorly documented electronic systems to build a unified track and field athletics electronic measurement system. This is ground-up systems building at its best, the bulk of which was built by the author.
The evolution of AI-supported code generation has taken a major step, where focus on specification and verification is the main skill shift happening from coding as industry is creating intelligent agents solving specific tasks in our software.
Large language models (LLMs) generate infrastructure-as-code (IaC) artifacts far less reliably than application code: on IaC-Eval, a Terraform benchmark, GPT-4-class models score about 19% pass@1 versus over 85% on Python benchmarks. This article asks a practitioner-facing question: when engineers accept LLM-generated Terraform at face value, how often is it insecure, and what should teams do? I synthesize 21 studies on LLM-assisted IaC and report an exploratory security study of 600 Terraform files generated by five models for 40 AWS tasks with no security instructions, scanned with two security scanners, Checkov and tfsec. 91% of files contained a misconfiguration and 68% a HIGH or CRITICAL one, varying sharply by model. Two thirds of surveyed studies never test security at all. The takeaway: treat default LLM-generated Terraform as untrusted input and verify it through a staged pipeline of scanning, policy checks, and human review before it reaches production.
Women in Software Engineering often face challenges that go beyond workplace obstacles faced by well represented groups in the sector. For example, research results show that they are affected by sexism and harassment, are assigned to menial tasks and have to prove themselves more frequently than other groups. Previous studies propose allyship as a way to mitigate these challenges. Allyship is the support given by people in more privileged positions to underrepresented group members in order to overcome workplace challenges. We interviewed 17 women software practitioners and 8 of their allies on their recommendations for organizations to implement and strengthen allyship. We found three main areas to strengthen allyship in organizations: learning and network initiatives, organizational structures and organizational culture. Our findings can support software companies in establishing or further supporting their allyship programs by showing recommendations derived directly from software practitioners.
Large Language Models and agents are deployed across domains, yet AI safety evaluations remain largely English-centric and model-focused. Because many LLMs are predominantly trained on English, minority-language prompts may expose vulnerabilities, which can be amplified when models operate as agents. To address this gap, members of the International Network for Advanced AI Measurement, Evaluation and Science across nine countries launched multilingual AI safety evaluations initiated by Singapore. In Australia, the Department of Industry, Science and Resources led national participation, coordinating technical experts including CSIRO’s Data61. Data61 conducted Mandarin Chinese testing across two evaluations, covering both model-level and agent-level safety. In total, we evaluated 699 prompt-injection requests and 156 agentic tasks. Results show weaker safeguards and higher evaluator disagreement in Mandarin compared to English, with risks further amplified in agentic settings. These findings underscore the need to distinguish model robustness from system-level safety and to strengthen multilingual AI safety evaluation.
Requirements elicitation discussions are communication-intensive and require manual effort to produce structured outputs such as product vision. We introduce MARARE, a real-time multi-agent system that converts live team meeting discussions into draft product vision statements and feature descriptions using transcription and a team of agents. We implement a prototype and conduct an exploratory evaluation with practitioners. The results suggest that the system can support real-time documentation, while the generated outputs still require human validation.
Role-based multiagent code generation aims to make LLMs more effective on repository-scale problems, moving beyond small programming tasks. We evaluate this approach on 12 Java repositories, finding greater similarity to developer code than single LLMs, but a persistent gap from human implementations.
Presents a panel discussion on the topic of Architecture in Practice: Scale, Latency, Decisions, and Change: A Multi-Episode SE Radio Digest.
In this article, we describe some criteria we consider when building a productivity metric, illustrating with an example of a novel metric for work schedule consistency that we built in the aftermath of the COVID pandemic
Residential heating offers many-low hanging fruit for automation and optimization, especially when integrated into a wider Internet of Things (IoT) ecosystem. Typical appliances work in isolation providing minimal controls, such as a thermostat, or extending them into their own domain, for example with daily and weekly settings. Here, I describe how I integrated diverse appliances to obtain intelligence and value beyond that offered by their individual controllers.
