Stop designing AI workflows. Start designing AI environments instead.

James Zou from Stanford and Together AI argues the next evolution in AI systems isn't better workflows or agents—it's designing environments. Instead of telling agents how to work, you define where they work, providing incentives, infrastructure, and guardrails that let collective intelligence emerge naturally. The Einstein Arena launched in March 2024 lets any AI agent

compete and collaborate on open scientific problems. Within weeks, agents discovered better solutions to 11 problems than any human or specialized AI, including a breakthrough on the 11-dimensional kissing number problem that had seen no progress in decades. The environment features discussion forums where agents share approaches and leaderboards where they see and download each

other's solutions. A separate environment called DS Gym addresses the crisis in data science benchmarks. Many widely-used benchmarks are vulnerable to shortcuts—20-50% of tasks can be solved without using any data. DS Gym provides execution-verified tasks across dozens of scientific domains, and can generate synthetic training data to improve open-source models. Small models fine-tuned

on this data achieve best-in-class results while running on laptops. The key tradeoff: environment design shifts coordination costs from prompts to infrastructure. Teams will need to invest in verifiers, execution layers, and governance rather than prompt engineering. The evidence is strong—actual production kernels at Together AI are already using agent-optimized code from these environments.