Agent Post-Training, Frontier Evals and Environments Research

Details

The Agent Post-Training team creates the frontier agents OpenAI ships to the world. The team trains models used in Codex, ChatGPT, the API, and other frontier products: persistent, proactive intelligence that can operate computers, collaborate with people and other agents, and expand what people and organizations can imagine, attempt, and achieve.

We define next-generation agent capabilities, build the training signals that teach those abilities, and run experiments that make them real. Work spans coding, tool use, computer use, multi-agent coordination, long-horizon execution, factuality, instruction following, calibrated reasoning, and taste. The team builds data, environments, graders, training methods, and feedback loops that shape OpenAI's next agents and carries capabilities through major training runs into products.

Responsibilities

• Create ambitious reinforcement learning (RL) environments to push models to their limits and to measure frontier model capabilities, skills, and behaviors.
• Develop methodologies for automatically exploring model behavior.
• Study the science of measurement, including scalability, reliability, and variance of evaluation methodology.
• Help steer training for large training runs and inform what goes into major model runs.
• Design scalable systems and processes to support continuous evaluation.
• Build self-improvement loops to automate model understanding.
• Collaborate with researchers, engineers, product teams, infrastructure teams, and safety/alignment partners to decide what should go into model runs and to ship improvements into products used by real users.
• Deliver high-agency research and engineering work that lands directly in frontier models.

Requirements

• Strong technical fundamentals in machine learning, software engineering, systems, statistics, or a related field; ability to learn quickly across parts not previously worked in.
• Hands-on experience with large language models (LLMs), reinforcement learning (RL), RLHF/RLAIF, post-training, evaluations (evals), graders, synthetic data, model training, coding agents, tool-using agents, or production ML systems.
• Experience designing experiments end-to-end: define hypotheses, build pipelines, run models, analyze results, and iterate.
• Comfort working across research, product, infrastructure, data, evals, and safety boundaries and communicating clearly with varied teams.
• Aptitude for open-ended problems where the path is unclear and signals are noisy; mix of research taste and engineering execution.
• Focus on product impact and model behavior (usefulness, reliability, honesty, tastefulness) rather than only benchmark movement.
• Willingness to build load-bearing systems and processes when required.

(Examples of prior open-sourced evaluations built by researchers in this role: GDPval, SWE-bench Verified, MLE-bench, PaperBench, and SWE-Lancer.)

Benefits

• Base pay range: $295,000 – $445,000 (base pay may vary by market location and experience). Total compensation may include equity and performance-related bonuses.
• Medical, dental, and vision insurance with employer contributions to Health Savings Accounts.
• Pre-tax accounts: Health FSA, Dependent Care FSA, commuter (parking and transit).
• 401(k) retirement plan with employer match.
• Paid parental leave (up to 24 weeks for birth parents, 20 weeks for non-birthing parents); paid medical and caregiver leave (up to 8 weeks).
• Paid time off: flexible PTO for exempt employees; up to 15 days annually for non-exempt employees.
• 13+ paid company holidays and multiple paid coordinated office closures; paid sick/safe time as required by law.
• Mental health and wellness support; employer-paid basic life and disability coverage.
• Annual learning and development stipend.
• Daily meals in offices and meal delivery credits as eligible.
• Relocation support for eligible employees.
• Additional taxable fringe benefits (charitable donation matching, wellness stipends) as applicable.