EurekAgent: A New Paradigm for Autonomous Scientific Discovery via Environmental Engineering

EurekAgent proposes a new research paradigm—Environmental Engineering—shifting the bottleneck of autonomous scientific discovery from agent workflow design to environment design. Through systematic design across four dimensions: Permission Engineering, Artifact Engineering, Budget Engineering, and Human-in-the-Loop Engineering, it has achieved new state-of-the-art results in mathematics, kernel engineering, and machine learning tasks, including discovering a new optimal solution to the 26-circle packing problem at a cost of less than $11.

In recent years, LLM agents have shown significant potential in automated scientific discovery, but the bottleneck has shifted from agent workflow design to environment design. The concept of Environmental Engineering was born from this need: by building and optimizing the agent's operating environment, it amplifies beneficial behaviors (such as exploration and collaboration) and suppresses harmful ones (such as reward cheating), enabling even simple agent architectures to produce effective behaviors.

The core idea of Environmental Engineering is that agent behavior is shaped by the environment. A well-designed environment should amplify beneficial behaviors (open exploration, systematic artifact management, collaboration) and suppress harmful ones (reward cheating, high-friction supervision). Unlike traditional approaches that focus on the agent's internal architecture, Environmental Engineering emphasizes the environment's role as a behavior shaper.

1. Permission Engineering: Precisely define execution boundaries to enable bounded execution and isolated evaluation, protecting system security while granting freedom for exploration;

2. Artifact Engineering: Manage knowledge assets using file systems and Git—each experiment corresponds to a Git repository, supporting multi-agent collaboration and process traceability;

3. Budget Engineering: Set budget limits, allowing agents to adjust strategies based on remaining budget to achieve low-cost breakthroughs (e.g., completing the 26-circle packing problem within $11);

4. Human-in-the-Loop Engineering: Design low-friction human-machine collaboration interfaces to support real-time intervention without disrupting experimental continuity, and record all interventions to ensure traceability.

EurekAgent has achieved results in mathematics (improving optimal packing problems), kernel engineering (more performant algorithms), and machine learning (new model architectures and strategies). Among these, the breakthrough in the 26-circle packing problem is the most notable: it found a more optimal configuration than previous solutions, with a total API cost of less than $11, verifying the effectiveness of Environmental Engineering.

EurekAgent marks an important turning point in the field of autonomous scientific discovery. The proposed Environmental Engineering paradigm provides a new thinking framework, promoting the application of LLM agents in a wider range of scientific and engineering fields. We look forward to more researchers making breakthroughs in this direction after its open-source release.

EurekAgent's implications: 1. Focus on the dimension of environment design; 2. Solve key challenges of autonomous systems (safety, reproducibility, cost-effectiveness, controllability) through engineering; 3. After the improvement of foundation model capabilities, system-level innovation becomes more important—future competition lies in transforming model capabilities into practical value.