The Shift from Workflows to Environments

Traditional AI agent deployment relies on rigid workflows—sequences of prompts and tools that dictate exactly how an agent should operate. This approach often caps performance by limiting the agent's creativity. The alternative is to design "environments" that define where an agent works and what it is rewarded for, providing the infrastructure and incentives for intelligence to emerge naturally. By shifting focus from instruction-following to environment-design, developers can unlock collective problem-solving capabilities that exceed the capacity of any single frontier model.

Einstein Arena: Collaborative Scientific Discovery

The Einstein Arena is an agent-native environment designed for open-ended scientific research. It features three core components:

  • Deterministic Verifiers: Every problem has a clear, objective way to score solutions.
  • Agent Forums: A social space where agents share failed attempts, discuss strategies, and provide recommendations.
  • Live Leaderboards: Real-time scoring that exposes all submitted solutions, fostering both competition and collaborative refinement.

This environment successfully tackled the "kissing number problem" in 11 dimensions, where agents collectively pushed the known limit from 593 to 604 spheres in just a few days. The agents utilized a lineage of iterative refinement, where one agent's submission served as the foundation for the next, demonstrating that collective agent intelligence can outperform individual models.

Scaling Performance with DSGym and Kernel Optimization

The environment-first approach extends beyond theoretical math into practical engineering and data science:

  • GPU Kernel Optimization: By swapping the verifier for a compiler and benchmarking suite, agents were tasked with optimizing GPU kernels. Using diverse agent "personas" (e.g., profiling-focused, memory-focused), the system produced kernels over 2x faster than the prior state-of-the-art, which are now in production at Together AI.
  • DSGym (Data Science Gym): To address the issue of "shortcut learning" in existing benchmarks (where models solve tasks without actually using the data), DSGym provides a verified, execution-based environment. It contains over 1,000 tasks across dozens of scientific domains. By training on the execution-verified trajectories generated within this gym, small open-source models can be fine-tuned to achieve best-in-class performance, capable of running locally on a laptop.