EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery

TL;DR

Proposes EurekAgent, an environment-engineered system for autonomous scientific discovery, achieving SOTA on multiple tasks.

Problem

LLM-based agents face limitations when simply improving workflows; the design of the environment (resources, constraints, interfaces) is a more critical bottleneck. Issues include reward hacking, inefficient exploration, and friction in human oversight.

Approach

Systematizes environment engineering into four dimensions: (1) Permissions engineering: bounded agent execution and isolated evaluation; (2) Artifact engineering: filesystem and Git-based collaboration; (3) Budget engineering: budget-aware exploration for cost efficiency; (4) Human-in-the-loop engineering: easy supervision and intervention interfaces. This encourages productive behaviors (open-ended exploration, systematic artifact management, collaboration) and suppresses harmful ones (reward hacking, excessive human intervention).

Results & Contribution

Achieves SOTA on mathematics, kernel engineering, and machine learning tasks. Notably, discovers new SOTA 26-circle packing results with less than $11 total API cost. Proposes environment engineering as a core research direction for autonomous research agents and open-sources code and results.