Coding agents have taken the world by storm. I will start this talk with a brief history of how we arrived here, beginning with the early days of using AI in software engineering. I will share experiences on working on AI for software engineering, and more specifically, on coding agents at two large companies. I will conclude by discussing how we can control agent behavior and what are some of the interesting problems in this space for software engineering researchers.

Satish Chandra is a research scientist at Meta, where he applies machine learning techniques to improve developer productivity. His work has spanned many areas of programming languages and software engineering, including program analysis, type systems, software synthesis, bug finding and repair, software testing and, of course, application of AI to software development. His research has been widely published in leading conferences in his field. Satish Chandra obtained a PhD from the University of Wisconsin-Madison, and a B.Tech from the Indian Institute of Technology-Kanpur, both in computer science. He is an ACM Fellow and an elected member of WG 2.4.

In recent years, Large Language Models (LLMs) have shown impressive performance across a wide range of downstream applications, including software engineering. This talk explores the evolution and recent trends of software engineering agents, featuring our recent work on Live-SWE-agent and Self-Play SWE-RL. We will discuss how these recent efforts enable live coding LLMs and software agents capable of autonomous and continuous self-improvement, paving the way toward next-level software intelligence.

Lingming Zhang is an associate professor at University of Illinois Urbana-Champaign. His research lies at the intersection of Software Engineering and Machine Learning. His group has pioneered a series of work on LLM-based software testing, repair, and synthesis (such as TitanFuzz, AlphaRepair, and Agentless), and also released multiple open code LLMs (including the recent SWE-RL, PurpCode, and Code World Model), with millions of downloads worldwide. Their techniques for training and improving code LLMs or agents have been adopted by leading AI companies, including Meta, Google, OpenAI, MiniMax, and DeepSeek.

What happens when AI agents leave the lab and land in the hands of real users? In this talk, I’ll share lessons from building and deploying agentic systems like GitHub Copilot and Excel’s M365 Copilot—where writing code is cheap, but testing and debugging are hard; prompts and tools age fast; and users don’t always know what to ask for. We’ll explore what’s working, what’s not, and why testing, trust, and thoughtful design matter more than ever. Along the way, I’ll offer a few critiques of our own assumptions—and a look at where we go from here.

Gustavo Soares is a Principal Research Manager at Microsoft, where he leads the Excel Agent Science team and drives research and product innovation at the intersection of agentic AI, large language models, and human–AI interaction. His work centers on the design, evaluation, and deployment of intelligent agents that can plan, reason, and act within complex productivity tools, with a strong emphasis on reliability, usability, and real‑world adoption at global scale. Previously, his research on agentic AI contributed to the development of GitHub Copilot agents for debugging and migration scenarios, and his work on modeless program synthesis helped lay the foundation for Visual Studio IntelliCode Suggestions, a capability used by millions of developers to automate repetitive and large‑scale code transformations. He received his Ph.D. in Computer Science from the Federal University of Campina Grande, Brazil; his doctoral thesis was recognized with the CAPES Thesis Award for the best Ph.D. thesis in Computer Science in Brazil, and he is also a recipient of the ACM John Vlissides Award.

AI-based and LLM-based systems are often cast as a threat to software engineering: coding agents grow more capable, junior roles shrink or disappear, and end-user programming appears ready to bypass experts altogether. This talk argues for a different conclusion. The deeper shift is not mainly about tools or tasks, but about organizational operating logic: how decisions are made, how knowledge moves, how work is coordinated, and how value is created. As these systems become more agentic, they will not merely support that shift; they will increasingly help analyze, adapt, and redesign it.

This change does not diminish the importance of software engineering. It broadens it. More of society will come to depend on semi-executable systems built from natural language, code, tools, policies, and workflows, while agentic environments increasingly generate, coordinate, and evolve many of those components. I take seriously concerns about reliability, maintainability, organizational inertia, and power. But these are better seen as constraints on the route than as arguments against the direction itself. Even imperfect AI can be transformative when deployed continuously at scale, and the central activities of software engineering are more likely to be redefined than replaced. For researchers, this creates a real risk that some AI4SE work targets practices already being folded into agentic workflows. For practitioners, it means that engineering discipline becomes more important, not less, as solution creation extends beyond traditional software developers.

Slides: Agentic Software Engineering Will Eat the World: AI-based Systems as the new OS of Society
Paper: The Semi-Executable Stack: Agentic Software Engineering and the Expanding Scope of SE

Robert Feldt is professor of software engineering at Chalmers University of Technology, Sweden, and at Blekinge Institute of Technology, Sweden. He has broad research interests spanning from human factors to hardcore automation and statistics, and including software testing and quality, requirements engineering, human-centred (behavioural) software engineering. He was an early contributor to search-based software engineering and have recently argued for increased application of psychology and social science to better understand and improve software engineering. Most of his research is empirical and conducted in close collaboration with industry partners in Sweden, Europe and Asia but he also leads more basic research. He received a PhD in computer engineering from the Chalmers University of Technology in 2002, studied psychology at Gothenburg University in the ’90s and has also worked as an IT and software consultant for more than 25 years. He is passionate about empirical research and methods and to change organisations through technological innovation.

AI coding agents are rapidly evolving and capable of performing complex software engineering tasks. As these systems move toward real-world deployment, rigorous and meaningful evaluation becomes increasingly critical for understanding their capabilities, limitations, and risks. In this talk, I present a series of evaluation methods that aim to bridge this gap, covering code generation, question answering and agentic systems. I will discuss lessons learned from building realistic evaluation frameworks and outline future directions toward trustworthy and effective AI software engineering agents.

Dr. Chao Peng is a Principal Research Scientist at ByteDance, where he leads the Software Engineering Lab focusing on AI agents for software engineering. His research interests include software testing, program repair and optimisation, as well as their synergy with machine learning and compiler techniques. His work has been published in premier venues such as ICSE, FSE, ASE, ACL, and NeurIPS. He received the Distinguished Reviewer Award at FSE 2025.