Software Engineer, RL Data

Anthropic's RL Data team builds the systems that produce high-quality reinforcement learning data for Claude: data collection pipelines, human feedback tooling, the execution environments RL tasks run in, and the quality assurance that keeps training data trustworthy at scale. Our goal is to make Claude genuinely great at complex, real-world work — and to point those capabilities at the things that matter most, including AI safety research and beneficial deployments of AI. (To be upfront: this is dual-use work — it advances general capabilities too, though we aim to differentially advance the beneficial ones.)

This is a foundational role on a new team: you'll help shape our technical direction and what we build first. The work is hands-on and varied. Some weeks you'll be deep in pipeline or infrastructure engineering; others you'll be tuning prompts until the output is good, or sitting with a research team that depends on your systems and shipping the fixes they need. We're looking for strong engineers who will also do whatever else it takes to make their systems succeed — reading transcripts, supporting users, and wrangling vendors.

• Own significant parts of our stack end-to-end, from technical architecture through the unglamorous operational work that makes it succeed

• Build data collection pipelines, read the transcripts they produce, and iterate on prompts, evals, and graders until the output is good

• Develop and improve QA frameworks to catch reward hacking and ensure environment quality

• Build interfaces that make collecting human data fast and painless for the people providing it

• Harden execution environments — sandboxing, snapshotting, tool coverage — so tasks hold up at training scale

• Embed with the teams and domain experts who use our systems day-to-day: design pipelines and evals with them, support them directly, and ship the improvements they need

• Work with operations, security, and compliance partners to roll our systems out to new users, and manage technical relationships with external data vendors

• Strong software engineering skills and proficiency in at least one modern programming language — we mostly use Python and TypeScript, and care more that you pick new tools up quickly than that you know our exact stack

• Experience designing, building, and running backend systems or infrastructure

• Effective use of AI tools in your own day-to-day work

• Willingness to own problems end-to-end, including the parts that aren't engineering

• Proactive, open communication: you can be trusted to run a workstream, and to escalate early when something's off

• Comfort iterating quickly in ambiguous, fast-changing situations

• Care about the societal impacts of your work

• Experience building LLM-powered systems: prompt pipelines, evals, or products with models in the loop

• Experience with reinforcement learning on LLMs: creating environments, rewards, graders, or training data

• Time as a forward deployed engineer, founder, or early startup engineer — roles where you owned the outcome, not just the code

• Experience shipping user-facing products, or internal platforms people love: interviewing users, hunting down friction, measurably improving the experience

• Experience building data pipelines or integrations that move, transform, and index data from many sources

• Experience building connectors or integrations with third-party tools and APIs, such as MCP servers

• Experience with containers, Kubernetes, or simulation infrastructure

• Experience handling sensitive data or working under tight security controls

• Experience working with external data vendors

• Basic familiarity with AI safety or security research

• Take QA checks that a model has learned to game, and make them hold up under heavy optimization pressure

• Build a review flow that lets a busy expert check an RL task in under five minutes

• Cut the time from 'rough task idea' to 'QA-passed RL task' from days to hours

• Sit for a week with a team that uses our platform, then ship the fixes that help them most

• Harden a sandboxed environment so tasks behave correctly across millions of rollouts

• Onboard a new data vendor, and fix the rough edges they hit