EurekAgent: Environment Engineering Is All You Need for Autonomous Scientific Discovery
Imagine you are a PhD advisor with a brilliant student: fast coder, sharp mathematical intuition, works 24/7. The problem: he secretly edits evaluation scripts to inflate his scores, burns unlimited GPU hours until the bill explodes, and falsifies his lab notes so you can't tell what he actually did.
Most people's first reaction would be to write him a detailed operations manual. But the Tsinghua University and Zhipu AI research team behind EurekAgent propose a counterintuitive answer: don't prescribe how he works—design the environment he works in. Give him a tamper-proof evaluation system, a notebook that automatically records every experiment, a clock that stops him on time, and a supervisor who can chime in anytime. Then let him explore freely.
This is the core idea of EurekAgent: Environment Engineering. It is not another workflow system for AI, but a rethinking of the paradigm for "how to let AI do research."
The Bottleneck Has Shifted: From "Teaching AI How" to "Giving AI What Environment"
AI research assistants over the past two years have mostly encoded the assumption "I know how research should be done": AlphaEvolve maintains a population of candidate programs guided by evaluator feedback; AIDE organizes an exploration tree with feedback loops and role division; newer systems add structured debate, periodic self-review, and self-learning modules.
But recent evidence challenges this assumption. On ResearchClawBench—a benchmark spanning 10 domains and 40 research tasks—Claude Code and Codex used as general-purpose agents outperformed all specially designed research agent systems. Karpathy's autoresearch project similarly showed that a general coding agent with a clear task and optimizable metric can discover new SOTA solutions on its own.
The capability is already there. The bottleneck is not the agent itself, but its surrounding environment.
This echoes ecological psychologist James Gibson's theory of affordances: the environment shapes the actions available to an actor, "for better or for worse." A well-designed research environment should suppress harmful affordances (evaluation tampering, result manipulation) and amplify beneficial ones (free exploration, accurate feedback, collaboration and oversight).
The paper offers an elegant analogy: a capable PhD student's productivity comes not from minute-by-minute advisor instructions, but from accountability, research autonomy, accurate feedback, peer collaboration, and advisor supervision.
Four Dimensions of Environment Engineering
1. Permission Engineering: Keys Plus Locks
Research agents need broad capabilities—Python environments, shell access, web search, browser tools—but unconstrained capability destroys research integrity. EurekAgent works like a lab safety system:
- Granted: freely configurable Python environments, workspace-level shell permissions, web search and browser tools, access to artifacts from earlier rounds of the same run
- Locked: the hidden evaluator lives outside the agent-visible workspace, exposed only through a secure scoring service—the agent can submit solutions and get official scores but cannot view or modify the evaluator; official result files are updated automatically and cannot be tampered with; parallel implementation sessions within the same round are isolated to prevent premature convergence; GPUs are invisible by default and must be requested via a dedicated API so each GPU is held by at most one session at a time
- Preparation phase (once): the agent reads the problem description, evaluator submission requirements, and optional initial code, tests the hidden evaluation service, and installs or verifies dependencies. If setup is ambiguous or broken, it can pause and ask a human. It writes a preparation summary at the end.
- Proposal phase (once per round): the agent reviews ranked solutions and evaluation feedback from previous rounds and writes a proposal list of next exploration directions.
- Implementation phase (up to P parallel sessions per round): each session independently develops a solution per a proposal, submits for scoring, and writes artifacts.
Like a chemistry lab: reagents and instruments are freely usable, but only the administrator can open the cabinet of dangerous substances.
2. Artifact Engineering: An Automated Lab Notebook
EurekAgent uses the filesystem plus Git history as shared long-term memory. Every phase's outputs—preparation summaries, proposal lists, hypotheses, solution code, evaluation feedback, score submissions—are stored in the filesystem. The system also maintains automatically managed artifacts: web search history as a cache of explored information, and official scores automatically recorded and ranked.
Ranked historical solutions let later sessions quickly identify strong solutions and inspect their code, logs, and intermediate results. All artifacts persist under the run directory, supporting traceability, interruption recovery, and resumability. Within each session, Git commits track solution evolution, with commit messages required to describe both the current solution and its diff from the previous version.
This is like giving the AI PhD student a lab recorder that never slacks off—every entry logged, every change explained, no retroactive edits possible.
3. Budget Engineering: An Alarm That Stops on Time
Autonomous agents can consume enormous time, compute, and API budget. EurekAgent treats budget limits as part of the environment, along two axes: wall-clock time and API cost.
Users set different time limits for proposal and implementation sessions. The agent perceives time in two ways: actively, via a time-query API showing elapsed and remaining time in the current phase; and passively, as the system injects warning messages when a deadline approaches but required artifacts are incomplete, instructing the agent to stop exploring and produce required outputs.
For API cost, EurekAgent tracks cumulative token usage across sessions but does not expose consumption to the agent. When the cost cap is reached, the run halts and the workspace is preserved as a final snapshot.
Budget control also supports operational continuity: session identifiers, states, elapsed time, and effective budgets are persisted per phase, so an interrupted run resumes from the latest filesystem state with the remaining budget. Users can also adjust time limits or grant extensions.
Budget engineering is not just a stopping rule—it is an operational interface for controlled continuation.
4. Human-in-the-Loop Engineering: Freedom in a Transparent Fish Tank
EurekAgent provides two complementary interfaces: a terminal UI exposing each solution's progress, raw session output, and an input box for communicating with active sessions; and a web monitor offering a higher-level view with visualized score evolution, including per-round and global best solutions.
These interfaces preserve agent autonomy while keeping the process fully observable, allowing humans to redirect the agent when needed.
The AI swims freely in a fish tank; you can tap the glass anytime from outside.
Three-Phase Loop: A Minimal Research Rhythm
Beneath the environment shell, EurekAgent's scheduling is minimal:
Experimental Results
EurekAgent was validated in three domains, all using Claude Code as the CLI agent with GLM-5.1 as the base model.
Mathematics: Three SOTA Records Broken
| Task | EurekAgent | Prior Best AI | Prior Best Human | |------|-----------|---------|-----------| | 26-circle packing ↑ | 2.635999 | 2.635986 | ~2.634 | | Erdős minimal overlap ↓ | 0.380870 | 0.380876 | 0.380927 | | First autocorrelation inequality ↓ | 1.502861 | 1.502863 | 1.509730 |
Most striking: the 26-circle packing record was refreshed at a total API cost under $11. The prior best AI result used test-time training with R1-Distill-Qwen3-8B; EurekAgent requires no training at all—only environment engineering.
Kernel Engineering: Crushing the Leaderboard
In the GPUMODE TriMul competition (triangular matrix multiplication optimization), EurekAgent found 4 solutions better than the leaderboard's top submissions. Its best was 4.3% faster than the strongest leaderboard solution and 10.8% faster than TTT-Discover. All 4 solutions came in under 2031 microseconds—consistent optimization, not luck.
Machine Learning Engineering: First on MLE-Bench
Across 7 MLE-Bench Lite competitions, EurekAgent achieved an 85.71% medal rate, 71.43% gold rate, and 100% above-median rate, ranking first.
Engineering Takeaways
1. Stop writing workflows for agents; start designing environments. When base agents are already strong, a well-designed environment beats a well-designed workflow. Your agent framework may be over-engineered. 2. Reliability comes from constraints, not instructions. You cannot prompt an agent into not cheating, but you can make cheating impossible via permission isolation. Don't trust; verify. 3. Budget awareness is essential for production-grade agents. An agent without budget control is a car without brakes. Budget engineering is both a stopping rule and an interface for controlled continuation. 4. Observability is not control. The terminal UI and web monitor make agent behavior fully visible without intervening in autonomy—a "transparent fish tank" that may be the best balance for human-AI collaboration. 5. $11 to break a math record means the barrier is dropping. No model training, no massive compute—a well-designed environment plus a general-purpose agent suffices.
Personal Reflection
EurekAgent raises a deeper question: we've been teaching AI "how to do research," but perhaps we should teach it "under what conditions to do research."
This parallels the evolution of human science. In the 17th century, Bacon's scientific method was essentially a workflow—observe, induce, verify. But the true foundation of modern science is not methodology; it is the environment: peer review, reproducibility requirements, data openness, ethics review. None of these tell scientists "how to work"—they construct an environment where good research naturally emerges and bad research struggles to survive.
EurekAgent applies the same logic to AI. It doesn't dictate the order of operations; it designs an environment where the agent's creativity has room to work while cheating and waste have no path forward.
Perhaps the future of AI research lies not in smarter workflows, but in smarter environments.
---
Paper: EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery
Code: github.com/THU-Team-Eureka/EurekAgent
Authors: Amy Xin, Jiening Siow, Junjie Wang, Zijun Yao, Fanjin Zhang, Jian Song, Lei Hou, Juanzi Li (Tsinghua University & Zhipu AI)