As language models have advanced, they have moved beyond code completion and are beginning to tackle software engineering tasks in a more autonomous, agentic way. However, evaluating agentic capabilities is challenging. To address this, we first introduce SWE-bench, a benchmark built from real GitHub issues that has become the standard for assessing AI’s ability to resolve complex software tasks in large codebases. We will discuss the current state of the field, the limitations of today’s models, and how far we still are from truly autonomous AI developers. Next, we will explore the fundamentals of agents based on SWE-agent, a simple yet powerful agent framework designed for software engineering but adaptable to a variety of domains. By the end of this talk, you will have an understanding of the current frontier of agentic AI in software engineering, the challenges ahead, and various tips and tricks on optimizing AI agents for tool use and iterative problem solving of reasoning-heavy tasks.

The open-source AI ecosystem is evolving rapidly, with frequent releases of new models, architectures, and hardware accelerators. While this flexibility drives innovation, it also introduces significant hidden complexities when fine-tuning and deploying AI models. In this talk, we’ll explore the key challenges teams face when updating fine-tuned models, switching between inference engines, and deploying across different GPUs such as the Nvidia A100, L40s, H100, and Intel Gaudi 2. We’ll share real-world examples, including tokenizer issues, multi-GPU fine-tuning hurdles, and API inconsistencies across AI components.

This talk explores the cutting-edge development of a multi-agent AI system designed for open-ended game environments, with a focus on Minecraft. The presenter will discuss the creation of specialized AI agents, including a Vision Agent for spatial analysis, a Curriculum Agent for dynamic task generation, an Action Agent for behavior execution, a Critic Agent for performance evaluation, and a Skill Manager for knowledge retention. The presentation will highlight how these agents work together, leveraging advanced techniques like Reinforcement Learning, Chain of Thought, and Tree of Thought reasoning to enhance decision-making and adaptability. A key feature of this project is the multi-bot system, which enables multiple AI agents to operate concurrently in the same game world, fostering collaboration and skill sharing. By demonstrating the potential of AI-driven automation and adaptability in complex, dynamic environments, this research opens up exciting possibilities for applications beyond gaming.