Taming GraalVM Reflection with AI Agents: Lessons from Testing 1000 Libraries
Conference (INTERMEDIATE level)
Auditorium
Making GraalVM work with the entire JVM ecosystem has historically meant one thing: manual, painful configuration of reflection metadata. We decided to automate this by generating comprehensive test suites for over 1000 key JVM libraries using AI and collecting the reflection metadata along the way. We learned that "asking the LLM" is not enough.
In this session, we dive deep into the architecture of an autonomous test generation pipeline. We move beyond simple prompts and demonstrate a feedback-loop architecture where GraalVM analysis and coverage data directly guide AI agents to uncover hidden code paths and edge cases.
You will learn:
Join us for a data-driven journey into the future of automated compatibility testing.
In this session, we dive deep into the architecture of an autonomous test generation pipeline. We move beyond simple prompts and demonstrate a feedback-loop architecture where GraalVM analysis and coverage data directly guide AI agents to uncover hidden code paths and edge cases.
You will learn:
- The Architecture: How to orchestrate agents to minimize token costs while maximizing coverage.
- The Feedback Loop: Techniques for injecting compile-time and runtime metrics back into the context window to increase coverage.
- The Benchmark Results: A transparent look at the cost (USD), time, and code coverage of different techniques and state-of-the-art models (GPT-5 vs. Open Source) when applied to real-world Java libraries.
Join us for a data-driven journey into the future of automated compatibility testing.
Vojin Jovanovic
Oracle
Vojin is an engineer at Oracle, where he leads projects in GraalVM developer experience and ML-based static profiling. Beyond his core work, he explores orchestration of AI agent teams to automate complex tasks. He holds a PhD in embedded domain-specific languages from the Scala Lab at EPFL.
Mihailo Markovic
Oracle
Mihailo is a Software Engineer at Oracle Database with a strong interest in machine learning, computer systems, and programming languages. He is currently developing infrastructure and workflows for the full automation of GraalVM Native Image compatibility across the Java ecosystem. As part of this work, he evaluates the performance of different techniques and design choices in AI automation, including context engineering, prompt design, and workflow architecture.
He has also contributed to research at INESC-ID in Lisbon, where he worked on a cloud-based runtime environment for WebAssembly modules that exposes registration and invocation through a REST interface, supporting serverless execution across multiple programming languages. Alongside his industry and research work, Mihailo serves as a Student Demonstrator at the University of Belgrade, School of Electrical Engineering, assisting in the evaluation of student work across multiple courses.
He has also contributed to research at INESC-ID in Lisbon, where he worked on a cloud-based runtime environment for WebAssembly modules that exposes registration and invocation through a REST interface, supporting serverless execution across multiple programming languages. Alongside his industry and research work, Mihailo serves as a Student Demonstrator at the University of Belgrade, School of Electrical Engineering, assisting in the evaluation of student work across multiple courses.