Member of Technical Staff, RL Research & Environments

Magic’s mission is to build safe AGI that accelerates humanity’s progress on the world’s most important problems. We believe the most promising path to safe AGI lies in automating research and code generation to improve models and solve alignment more reliably than humans can alone. Our approach combines frontier-scale pre-training, domain-specific RL, ultra-long context, and inference-time compute to achieve this goal.

ABOUT THE ROLE

As a Software Engineer on the RL Research & Environments team, you will design and operate the data, evaluation, and environment systems that improve model capabilities after pre-training.

This role focuses on post-training: identifying capability gaps, building targeted datasets, designing reward signals, and running iterative training loops that measurably improve user-facing behavior. You will own the infrastructure and experimental workflows that connect product priorities to concrete capability gains.

Magic’s long-context models introduce distinct post-training challenges: long-horizon reasoning, sustained coherence over extended trajectories, context-use quality, and tool-augmented behavior. You will build systems that expose failure modes, generate high-signal training data, and enable rapid RL iteration at scale.

This role can evolve into ownership of major capability areas, deeper RL systems work, or broader influence over post-training strategy as Magic scales long-context model performance and reliability.

WHAT YOU’LL WORK ON

- Design and build post-training datasets using synthetic generation, targeted data collection, and self-play

- Implement filtering, scoring, and mixture strategies for RL and post-training corpora

- Build and maintain evaluation frameworks that surface long-context failure modes

- Design reward signals and training environments for targeted capability improvements

- Run ablations across data sources, reward designs, and long-horizon task structures

- Improve reliability and o...