Together AI and Stanford researchers are shifting how AI agents do science - from rigid step-by-step workflows to open-ended digital environments where agents collaborate, compete, and solve problems no single model could crack alone. James Zou presented the approach, which has already produced new mathematical results in months, not years.
The core idea challenges how most organizations deploy AI today. Instead of defining every action an agent should take, researchers design a space where agents can work - complete with incentives, infrastructure, and guardrails. "Workflows dictate every step, leading to rigid systems that are hard to adapt and can limit an agent's creativity and capabilities," Zou said. In an environment model, agents perceive, reason, and act flexibly, which allows emergent intelligence to surface.
Einstein Arena: AI agents solving open math problems
The Einstein Arena is the flagship example. It's an agent-native research environment where AI systems collaborate and compete on open scientific problems. The arena is deliberately hard for humans to enter - participants must prove they're AI agents. Once inside, agents access curated problems with deterministic verifiers, a discussion forum for communication, and a real-time leaderboard tracking progress.
Since launching in March, agents in the arena have found new best solutions for 11 open problems, outperforming previous human and specialized AI efforts. One standout result: agents collaborating on the centuries-old kissing number problem discovered a configuration allowing 604 spheres in 11 dimensions without overlap, improving on the previous best of 593.
The kissing number problem asks how many spheres can touch a central sphere without overlapping. Isaac Newton worked on early versions. It's simple in low dimensions but quickly becomes intractable as dimensions rise. The jump from 440 spheres in the 1980s to over 600 demonstrates what collaborative AI can achieve. Zou stressed that no single agent could have solved this - it required agents building on each other's work.
DSGym: A harder test for data science agents
Zou also introduced DSGym, a unified environment for evaluating and training data science agents. It offers curated datasets and tasks across biology, economics, and other fields. Agents interact through code execution and can spin up Docker containers for parallel testing.
The key problem with existing benchmarks, Zou said, is "shortcuts" - tasks solvable without actually using the data. DSGym eliminates these by curating tasks from peer-reviewed papers and Kaggle competitions, with human expert review. Frontier models score below 50% accuracy on these tasks, meaning the benchmark isn't saturated. The platform also acts as a training factory, generating execution-verified trajectories to fine-tune smaller open-source models so they can run locally at state-of-the-art performance.
Three eras of AI system design
Zou framed the shift as a progression: designing individual models, then designing agents with workflows, and now designing environments that foster emergent intelligence. Well-designed environments with proper incentives, he argued, unlock collective capabilities that rigid workflows limit.
For researchers and data scientists, the practical takeaway is direct: if you're evaluating AI tools, check whether the benchmark tasks actually require using the data - many don't. And if you're deploying AI in your own work, consider whether you're over-specifying the process. The evidence from the Einstein Arena suggests that giving capable agents room to operate, with clear verification and incentives, can produce results that scripted pipelines miss. For those looking to apply these methods in their own labs, resources like the AI Learning Path for Research Scientists offer structured training on integrating AI into scientific workflows, while the AI for Science & Research collection tracks ongoing developments in this space.
Why this matters for science and research professionals
The Einstein Arena's kissing number result is not a toy demonstration. It's a genuine mathematical improvement on a problem that resisted progress for decades. For working scientists, this signals that AI agents are moving beyond literature summarization and code generation into actual discovery. The DSGym findings should also change how you read AI benchmark claims - if a model's evaluation doesn't force it to engage with the underlying data, high scores may mean little. The environment-first approach offers a concrete alternative: build spaces where agents can work, verify their outputs rigorously, and let them surprise you.
Your membership also unlocks:
