Research Engineer / Performance Engineer, RL Distributed Systems

Reinforcement learning is how Claude learns to reason, write code, and act autonomously over long horizons. At frontier scale, an RL run is an unusually demanding distributed system. Training, sampling, and environment execution run concurrently across a large fleet of accelerators and hosts, exchange data continuously, and have to keep making progress while hardware fails, load shifts, and the research changes underneath them. How well that system holds together determines how much of our compute turns into learning, and how quickly the team can try the next idea.

As a Research Engineer on the Distributed Systems team within RL Engineering, you'll work on whatever part of that system is the current limit. That might be scheduling and placement, data movement between components, running large numbers of sandboxed environments, storage and checkpointing, networking, fault tolerance, autoscaling, or the observability that tells us what a run is actually doing. We're looking for generalists: engineers who can move between these layers, reason from first principles about a system they haven't seen before, and pick the problem that matters most rather than the one closest to their prior experience.

Our system changes as fast as the research does, correctness under failure matters as much as throughput, and the best solutions often come from understanding the ML workload well enough to know which guarantees it actually needs. Strong candidates have built and run large distributed systems, care about getting the details right, and want to apply that experience to a workload that is very large, very heterogeneous, and changing quickly.

• Design, build, and operate the distributed systems that run RL at scale, across training, sampling, and environment execution
• Find and remove whatever currently limits the system, whether it's scheduling, data movement, storage, networking, or coordination
• Build fault tolerance into every layer: failure detection, isolation, and recovery that keep long-running jobs making progress without human intervention
• Design resource management and autoscaling so that compute follows demand as a run's needs shift
• Build observability that makes it possible to understand what a run is doing and why it slowed down, stalled, or produced unexpected results
• Build automation that detects and remediates common problems, and design interfaces that let engineers and automated tools operate runs safely
• Work with researchers and performance engineers to make sure systems changes preserve training correctness and don't introduce subtle nondeterminism
• Remove classes of failure at their source through incident review, testing, and redesign, and write clear design documents for what you build

• Strong software engineering skills in Python and at least one systems language such as Rust, C++, or Go
• Experience designing, building, and operating large-scale distributed systems in production
• Deep understanding of distributed systems fundamentals, including consistency, coordination, consensus, failure modes, and recovery
• Ability to reason quantitatively about throughput, latency, and resource costs across compute, memory, storage, and network
• Experience debugging complex failures across many hosts and services, including failures you can't reproduce locally
• Strong written communication, including design documents and incident writeups

• Experience running ML training or inference infrastructure at scale
• Experience across several layers of the stack, such as scheduling, storage, networking, and orchestration
• Experience building schedulers, autoscalers, or resource management systems
• Experience with container orchestration such as Kubernetes, and with sandboxed or virtualized code execution at scale
• Experience with high-performance networking, RDMA, or collective communication libraries
• Experience building observability or automated remediation for large fleets
• Experience with async Python frameworks such as Trio or asyncio
• Familiarity with reinforcement learning or large language model training workloads

• Design a scheduler that places training, sampling, and environment work across a heterogeneous cluster while respecting network topology and failure domains
• Build a failure detection and recovery system that lets a long-running job survive host and network failures with minimal lost work
• Scale environment execution substantially without increasing tail latency for the training step
• Design an autoscaling policy that rebalances compute across components as a run's bottleneck shifts
• Build a diagnostics system that explains why a run's throughput dropped and proposes a fix
• Trace a rare data corruption bug across many services to a race condition in a recovery path, and redesign the path so the class of bug can't recur
• Design the operational interface for a run so that automated tools can safely diagnose and adjust it under human oversight