About the Role The Computer-Use Agent (CUA) team at Hark builds agents that use a computer effectively and efficiently to accomplish increasingly long-horizon tasks. These are the models that let Hark get real work done on a user's behalf, from a single errand to workflows that run for hours.
We're looking for engineers and researchers who are motivated to build the best computer-use agents in the world. This role demands high-quality engineering combined with genuine research ability, and the agency to investigate promising novel algorithms and approaches that will shape the next generation of computer-use agents at Hark.
You'll work collaboratively in a fast-moving, high-performance team across every aspect of training computer-use agents. We expect deep technical expertise in one or more of the following areas: data and environment engineering, agentic evals, RL post-training, and training and inference infrastructure for agents.
This is a high-ownership role on a small team. Beyond the CUA team, you'll work closely with the harness and product, pretraining, and infrastructure teams.
ResponsibilitiesThe ideal candidate brings exceptional depth in one or more of the areas below, paired with a working understanding of the rest.
- Environments & Data: Build and scale interactive environments and task generators for computer use, including realistic applications, stateful workflows, verifiable success checks, and synthetic trajectories for SFT and RL.
- Evaluation: Design benchmarks and eval harnesses that measure task success, reliability, efficiency, and safety on long-horizon tasks, and turn model failures into concrete data, environment, or training fixes.
- Post-Training: Develop and scale SFT, RL (PPO/GRPO and beyond), reward modeling, and grader and verifier techniques that make agents both more capable and more efficient, including in long-horizon and sparse-reward settings.
- Research & Novel Methods: Identify promising new algorithms and approaches, prototype them quickly, and test them rigorously with strong baselines and ablations to decide what goes into the next major model run.
- Data Quality: Build filtering, curation, difficulty calibration, and curriculum methods so every training run gets the highest-signal data we can produce.
- Agent Infrastructure: Build and scale the training and inference infrastructure behind large-scale agent rollouts and RL, improving throughput, latency, reproducibility, observability, and cost.
- Production Impact: Partner with harness, product, and infrastructure teams to close the loop between real user tasks and model improvements, and debug hard failures in shipped models.
Requirements- Engineering Excellence: Strong software engineering and prototyping skills, with a track record of building high-quality, reliable systems that other researchers and engineers depend on.
- Research Ability: Demonstrated ability to turn an open-ended problem into a crisp hypothesis, design rigorous experiments, and reason clearly from noisy results.
- Model Experience: Hands-on experience training, fine-tuning, or evaluating large language or multimodal models.
- Domain Depth: Exceptional ability in one or more of the responsibilities above, with deep expertise in at least one of data and environment engineering, agentic evals, RL post-training, or training and inference infrastructure for agents.
- Agency & Ownership: Comfort with ambiguity and a bias for action. You chase promising ideas independently and do whatever the problem needs, including unglamorous, load-bearing work.
- Communication: Clear technical communication and the ability to collaborate across research, infrastructure, and product.
Bonus Qualifications- Experience training models for computer use, coding agents, web or browser agents, or other tool-using agents.
- Experience with RL in long-horizon or sparse-reward settings, or with reward models, LLM judges, and verifiers.
- Experience with large-scale distributed training or rollout systems (e.g., Ray, sandboxed VMs or containers, browser automation).
- Publications or open-source work in agents, post-training, evaluation, or multimodal learning.
- Experience translating product signal into model improvements for a shipped AI product.
CompensationThe US base salary range for this full-time position is between $180,000 - $400,000 annually.
The pay offered for this position may vary based on several individual factors, including job-related knowledge, skills, and experience. The total compensation package may also include additional components/benefits depending on the specific role. This information will be shared if an employment offer is extended.
