RL Environments Architect: Build Safe, Scalable Agents

RL Environments Architect; United States; Remote; USD 180,000 - 240,000; Full time

Job summary

Surge is seeking an RL Environments Architect to design, instrument, and govern the simulated worlds where agents learn. You’ll define primitives, reward structures, interfaces, and telemetry to stress-test capabilities while keeping signals faithful, stable, and scalable.

You’ll also establish data quality standards and reproducibility across large-scale agent gyms, turning messy real-world phenomena into learnable curricula and building the backbone for safe, rapid progress in agentic systems.

Qualifications

• Experience building RL environments or simulators with an eye for determinism, performance, and observability.
• Strong instincts for designing reward functions, scenario taxonomies, and QA pipelines that keep signals aligned and drift-free.
• Collaborative mindset across research and engineering to ship pragmatic, testable environments that evolve with model capabilities.

Responsibilities

• Architect a modular environment framework with clear APIs, curriculum scaffolds, and configurable reward/termination schemas
• Establish quality bars: coverage metrics, invariance checks, and trace audits for environment outputs and agent experience buffers
• Instrument rich telemetry for episode rollouts; mitigating reward hacking, mode collapse, and exploitable loopholes
• Partner with researchers to translate real-world tasks into robust simulations, including synthetic data generators and evaluation suites

Skills

Simulation & Systems Depth

Data Quality Leadership

Builder’s Mindset

Job description

Surge is seeking an RL Environments Architect to design, instrument, and govern the simulated worlds where agents learn. You’ll define primitives, reward structures, interfaces, and telemetry to stress-test capabilities while keeping signals faithful, stable, and scalable.

You’ll also establish data quality standards and reproducibility across large-scale agent gyms, turning messy real-world phenomena into learnable curricula and building the backbone for safe, rapid progress in agentic systems.