Research Scientist

Centaur is building agents that take ownership of complex objectives and pursue them autonomously for days, months, and eventually indefinitely, a new class of AI: unbounded, long-running agents that don't stop after a single task but keep making progress toward an outcome over time. Its founding team includes leaders and founding members from Tesla AI, Physical Intelligence, Anthropic, Adept, Jane Street, NVIDIA Research, and Google DeepMind.

The core thesis: as agents operate over longer horizons, the bottleneck shifts from capability to intent, accurately understanding what a person or organization wants, preserving that understanding over time, and knowing when human judgment is actually required. Centaur builds systems that turn sparse human direction and messy longitudinal evidence into a persistent representation of intent, then use it to plan, act, evaluate outcomes, recover from mistakes, and continuously improve.

In a world where human judgment is the scarce resource, the defining systems of this era will maximize valuable output per unit of human judgment at the lowest cost. Centaur's first application is self-improving products: agents that learn what customers are trying to accomplish, identify unmet and latent needs, build or configure solutions, deploy them, measure whether they worked, and keep iterating.

You'll train agents that use computers and tools to pursue complex outcomes over long horizons, owning ambitious research bets end to end, from hypothesis and data through training, evaluation, and deployment. The work centers on post-training LLM and multimodal agents with reinforcement learning and related methods, and on long-horizon agent capabilities: planning, memory, recovery, proactivity, self-improvement, and calibrated escalation.

A core thread is developing methods for agents to understand and preserve user intent across long-running trajectories. This is a build-real-systems research seat, not isolated prototypes: you'll build computer-use and tool-use environments, training-data pipelines, benchmarks, and evaluations; design rigorous experiments; identify the highest-use research problems; and turn successful results into deployed capabilities.

As an early member, you'll help shape Centaur's research agenda and technical direction. The bar is exceptional ML research and engineering, deep experience in one or more of the relevant areas, and comfort pursuing uncertain directions where the right approach isn't yet known.