Member of Technical Staff, Computer-Use Agents

Hark • San Jose, CA

11d • $180,000 - $400,000

About The Position

Hark is an artificial intelligence company developing proactive, multimodal, personalized intelligence that interacts with the world through speech, text, vision, and persistent memory. It is pairing this intelligence with next-generation hardware to build a universal interface between people and machines.

The Computer-Use Agent (CUA) team builds agents that use computers effectively and efficiently to complete increasingly long-horizon tasks, ranging from individual errands to workflows lasting hours. This role combines high-quality engineering with research ability and independent investigation of algorithms and approaches for the next generation of computer-use agents. The successful candidate will work across the training lifecycle with a collaborative, fast-moving team and collaborate closely with the harness and product, pretraining, and infrastructure teams. It is a high-ownership role on a small team, requiring deep expertise in at least one of data and environment engineering, agentic evaluations, RL post-training, or agent training and inference infrastructure.

Requirements

• Strong software engineering and prototyping skills, with a record of building high-quality, reliable systems used by researchers and engineers.
• Demonstrated ability to formulate crisp hypotheses from open-ended problems, design rigorous experiments, and reason clearly from noisy results.
• Hands-on experience training, fine-tuning, or evaluating large language or multimodal models.
• Deep expertise in at least one of data and environment engineering, agentic evaluations, RL post-training, or training and inference infrastructure for agents.
• Comfort with ambiguity, independent initiative, and willingness to take ownership of necessary work, including unglamorous, load-bearing work.
• Clear technical communication and ability to collaborate across research, infrastructure, and product.

Nice To Haves

• 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, such as Ray, sandboxed VMs or containers, or 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.

Responsibilities

• Build and scale interactive computer-use environments and task generators, including realistic applications, stateful workflows, verifiable success checks, and synthetic trajectories for SFT and RL.
• Design benchmarks and evaluation harnesses to measure task success, reliability, efficiency, and safety on long-horizon tasks, and translate model failures into concrete data, environment, or training fixes.
• Develop and scale SFT, RL (including PPO/GRPO and beyond), reward modeling, and grader and verifier techniques to improve agent capability and efficiency, including in long-horizon and sparse-reward settings.
• Identify promising algorithms and approaches, prototype them, and evaluate them rigorously using strong baselines and ablations.
• Build filtering, curation, difficulty calibration, and curriculum methods to improve the signal quality of training data.
• Build and scale training and inference infrastructure for large-scale agent rollouts and RL, improving throughput, latency, reproducibility, observability, and cost.
• Partner with harness, product, and infrastructure teams to connect real user tasks with model improvements and debug difficult failures in shipped models.

What This Job Offers

• Job Type: Full-time
• Education Level: No Education Listed