Researcher, Context - Agent Post-Training

OpenAI•San Francisco, CA

8d•$250,000 - $380,000

About The Position

The Agent Post-Training team at OpenAI is responsible for creating the frontier agents that OpenAI ships to the world. This includes training the models behind products like Codex, ChatGPT, and the API, focusing on developing persistent, proactive intelligence capable of operating computers, collaborating with humans and other agents, and expanding human potential. The team defines future agent capabilities, builds the training signals, and conducts experiments to realize these advancements. Their work encompasses areas such as coding, tool use, computer operation, multi-agent coordination, long-horizon execution, factuality, instruction following, calibrated reasoning, and taste. This team is where new model capabilities are developed, involving the creation of data, environments, graders, training methods, and feedback loops that shape OpenAI's next generation of agents, which are then carried through major training runs and into user-facing products. As a Context Researcher on the Agent Post-Training team, the primary focus is on scaling the compute spent on context, which is believed to be the final enabler for AGI. This role involves working within OpenAI's frontier training stack to enable the next paradigm of model training, with a clear product interface for iterative deployment (Codex Chronicle). The researcher will collaborate with various teams, including researchers, engineers, product teams, infrastructure teams, and safety/alignment partners, to determine the content of major model runs, measure their success, and implement improvements into widely used products. This is a high-agency position for individuals who want their work to directly influence frontier models.

Requirements

Strong technical fundamentals in machine learning, software engineering, systems, statistics, or a related field, and can learn quickly across the parts you have not worked in before.
Hands-on experience with LLMs, RL, RLHF/RLAIF, post-training, evals, graders, synthetic data, model training, coding agents, tool-using agents, or production ML systems.
Excited by open-ended problems where the path is unclear, the signal is noisy, and the right answer requires both research taste and engineering execution.
Care about product impact and model behavior, not just benchmark movement. You have opinions about what makes an agent useful, reliable, honest, tasteful, and easy to work with.
Can move from a vague behavioral problem to a concrete experiment: define the hypothesis, build the pipeline, run the model, analyze the result, and decide what to do next.
Comfortable working across research, product, infrastructure, data, evals, and safety boundaries, and can communicate clearly with each group.
Like building load-bearing systems and processes when that is what the team needs, even if the work is not glamorous.
Want to train and ship the models that make agents genuinely useful for developers, enterprises, researchers, and everyday users.

Responsibilities

Design and run experiments that improve scaling of compute on context.
Own end-to-end improvements to the post-training stack, including RL, data pipelines, graders, reward signals, evals, diagnostics, and model-behavior analysis.
Build evals and environments that expose the next set of model failures, then turn those failures into training data, product fixes, or new research directions.
Partner with Codex and ChatGPT product teams to understand what users need and translate product signal into model improvements.
Work on early-training and alignment interventions, including data mixtures, objectives, synthetic data, and eval loops that shape downstream agent behavior.
Help decide which integrations, capabilities, and fixes are ready for inclusion in major model runs.
Improve the machinery for large-scale training and launch: experiment velocity, reliability, observability, reproducibility, cost, latency, and production readiness.
Take on cross-functional projects that touch model training, product infrastructure, and the production agent harness, such as multi-agent systems or training directly against production-like environments.
Debug hard failures in shipped or near-shipped models and turn messy qualitative behavior into concrete hypotheses, experiments, and fixes.