> Markdown version of [/jobs/ext/4180863-senior-software-engineer-rl-environments-sandboxes](https://www.wearedevelopers.com/jobs/ext/4180863-senior-software-engineer-rl-environments-sandboxes). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # Senior Software Engineer - RL Environments & Sandboxes - **Company:** Lumos AI Research Corp. - **Location:** San Francisco, CA, United States - **Experience:** Expert - **Salary:** $160,000.0 - $220,000.0 - **Contract:** Permanent contract - **Skills:** Application Programming Interfaces (APIs), Artificial Intelligence, Software Applications, Computational Biology, Computer Programming, Data Structures, Software Debugging, Linux, Distributed Systems, Hardware Interface Design, Python (Programming Language), Open Source Technology, Software Architecture, Cyber-physical Systems, Scientific Computating, Software Engineering, Systems Integration, Reinforcement Learning, Model Validation, Backend, Agentic-AI, Information Technology, Docker - **Published:** October 8, 2026 - **Apply:** https://www.indeed.com/viewjob?jk=a0821637148d9d52 ## About the Role * 5+ years of software engineering experience, or equivalent evidence of exceptional engineering ability. * Excellent Python skills and strong fundamentals in algorithms, data structures, systems design, and software architecture. * Hands-on experience with Docker, Linux, containers, process isolation, or sandbox execution. * Strong understanding of distributed systems, asynchronous execution, state management, and reliability. * Experience designing complex APIs, backend systems, and integrations. * Ability to independently architect, build, debug, and ship sophisticated technical systems. We'd be particularly excited if you have * Hands-on experience with Harbor, especially building custom RL environments, task definitions, verifiers, or evaluation infrastructure. * Experience with RL environments, AI agents, model evaluation, or agent training infrastructure. * Familiarity with MCP, Gymnasium, Ray, or similar agent and execution frameworks. * Experience with robotics, laboratory automation, hardware interfaces, or cyber-physical systems. * Knowledge of deterministic simulation, sandbox orchestration, reproducible execution, and reward verification. * Experience with scientific computing, bioinformatics, or computational biology. The kind of person who thrives here We have a strong preference for intellectually curious, independent builders who enjoy unusually difficult problems and have a track record of pursuing ambitious ideas beyond what's expected of them. Some backgrounds we're especially drawn to: * Former founders: People who have built companies or products from scratch, are comfortable with ambiguity, and naturally take ownership. * Quantitative trading and low-latency engineering: Engineers from quantitative trading firms or similarly demanding environments who have developed exceptional problem-solving skills, strong systems intuition, and a high bar for correctness. * Competitive mathematics and programming: People with backgrounds in IMO, IOI, ICPC, or similar competitions who genuinely enjoy difficult technical and mathematical problems. * Computer science and philosophy: People who combine strong technical reasoning with an interest in logic, epistemology, formal systems, and foundational questions about intelligence. * Makers and tinkerers: People who build robots, hardware, electronics, open-source tools, or ambitious hobby projects simply because they find them interesting. These aren't requirements or substitutes for demonstrated engineering ability. They're examples of the curiosity, intellectual intensity, and builder mentality we value. ## Description You'll design and implement stateful, executable environments that allow agents to interact with software, APIs, scientific tools, and laboratory systems. These environments must be reproducible, verifiable, and scalable enough to support reinforcement learning and rigorous evaluation. This role is particularly exciting because your work won't be limited to simulated software. You'll work closely with our wet-lab team to build the infrastructure connecting AI agents to physical experiments, laboratory automation, and real biological outcomes. You'll own systems end to end, from environment architecture and sandbox execution to tool interfaces, task verification, and rollout infrastructure. We're looking for someone who enjoys working at the boundary of software, AI, and the physical world - someone who would be equally excited about building a distributed execution engine and making a laboratory robot perform an experiment. What Will You Do * Build RL environments: Design and implement realistic, stateful environments that reproduce complex workflows across healthcare, biology, and scientific software. * Develop secure sandboxes: Build isolated, reproducible execution environments supporting agent-generated code, tool calls, filesystem operations, and application interactions. * Connect AI agents to the wet lab: Work with scientists and automation engineers to integrate laboratory instruments, robotic liquid handlers, experimental protocols, and biological readouts into agent-accessible workflows. * Build agent-facing tools: Implement MCP servers, APIs, and interfaces that enable agents to interact with software applications, scientific instruments, and simulated systems. * Engineer state management: Develop deterministic initialization, snapshot/restore, state transitions, resets, and execution replay. * Develop tasks and verifiers: Translate real-world workflows into executable tasks with clear success criteria, ground-truth states, and programmatic verification. * Scale environment execution: Build orchestration infrastructure for high-throughput parallel rollouts, supporting both simulated environments and physical experiments with appropriate execution controls. * Work with Harbor: Develop, integrate, and validate environments using Harbor-compatible task structures, execution harnesses, and evaluation workflows. * Improve environment fidelity: Identify unrealistic shortcuts, hidden state inconsistencies, and verification gaps that could compromise training quality. * Own the full lifecycle: Take ambitious technical ideas from early prototypes through production-grade infrastructure.