Tartu — Workshop: AI Agents in Practice

The workshop introduces AI agents and how using them differs from having an ordinary conversation with a large language model. For example, an agent can work on a task for longer and turn ideas into prototypes. In the first half of the workshop, participants will learn how a large language model works and when a system can be called an agent. They will learn to distinguish between planning and building workflows and explore the role of tools and skills. They will discuss why vague instructions produce mediocre results, how to check outputs, and the risks involving data, copyright, and security. Participants will then create a plan based on their own idea or an assigned task (goal, input, output, quality criteria, and risks), build a prototype with an agent, and test it. After a break, they will present their prototypes and describe the agent’s errors and where human oversight is needed. Each participant will turn their result into a digital microlearning resource. Participants can then ask questions about skills needed in the labor market, using agents safely, and more. The workshop will end with a feedback round.

No programming experience is required, but curiosity and a willingness to experiment are welcome. Participants need a laptop, must create a GitHub account, and should complete the setup beforehand using the instructions. Tools used: OpenCode 1.19.4 (free and open-source agent software), the GitHub Codespaces cloud environment, and free models from OpenCode Zen. The work takes place in a cloud environment in a browser: nothing is installed on the computer, and the agent cannot access participants’ files. We use free tools and language models. Because text entered is sent to the model provider, participants are asked not to enter personal data or other sensitive information.

The workshop will be led by Heili Aavola, a junior researcher in machine learning at the University of Tartu’s Institute of Computer Science. Her doctoral research concerns uncertainty management in hybrid AI systems—in other words, whether an AI system can recognize when it might be wrong. Together with Martin Leissoo, she built an AI agent (Hermes) at the University of Tartu that runs on the university’s own hardware. She has a bachelor’s degree in chemistry and a master’s degree in data science. Before moving into data science, she worked in chemistry and microfluidics. The event is free.