The role
Turn questions about agent learning into reliable systems and reproducible experiments. Work directly with the founders to build environments, datasets, and evaluations, and study how task design, rewards, and verification affect model capabilities across long-horizon coding, AI4Science, AI4AI, knowledge work, and agentic tool use.
What you’ll work on
- Build realistic multi-step tasks, isolated runtimes, reference solutions, and verifiers, including environments in Harbor format.
- Develop pipelines for high-quality data generation, environment orchestration, agent rollouts, evaluation, and quality control at scale.
- Run controlled experiments on task difficulty, curricula, reward design, and long-horizon agent behavior.
- Analyze failures and reward exploitation, and build evaluations that measure generalization.
- Work with domain experts to turn scientific and knowledge-work workflows into reproducible training environments.
What you bring
- 3+ years of professional software engineering experience.
- Strong Python and experience with containers, data pipelines, and reliable experiment infrastructure.
- Hands-on work with reinforcement learning, model post-training, or agent evaluation; experience with PyTorch or a similar framework.
- Careful experimental judgment: clear baselines, controlled comparisons, and useful failure analysis.
- A track record of using coding agents to build, test, and debug substantial software.
Show us your work
Share a system, experiment, benchmark, paper, or open-source contribution you’re proud of. Explain what you built, what failed, and what you learned.