What is eurekagent?

EurekAgent is a research system from Tsinghua University that tries to automate scientific problem-solving using AI agents. You describe a problem, provide a way to measure how good a solution is, and EurekAgent runs a loop: it has AI agents propose approaches, write code to try them, execute the experiments in an isolated environment, score the results, and then use what it learned to try again. The process repeats until the budget or time runs out. The intended use case is computational research tasks where the goal is measurable -- things like optimization problems, algorithm design challenges, or any task where you can write a function that takes a proposed solution and returns a score. You define the problem in a plain text file, specify what format solutions should take, and provide a private scoring script. EurekAgent reads those files and treats the scoring script as the final authority on whether a solution improved. Each experiment runs inside a Docker container, which keeps the agent's code isolated from the host machine and from the grading process. Runs can be paused and resumed from their last saved state, which matters for long experiments. While a run is active, a web dashboard shows live costs, score history, and logs. The system is built on top of Claude Code, an AI coding tool from Anthropic, and is designed to work with several AI model providers including GLM-5.1. Optional browser access via a Playwright integration lets the agent search the web for relevant information during a run. Setting up a new problem can be done interactively: a built-in skill in Claude Code walks you through the setup and generates the required files from a plain-language description of what you want to solve. The repository includes a circle packing example that can be run immediately after installation to verify everything works.