Applied Research - Computer Use

Prime Intellect•San Francisco, CA

1d•$150,000 - $300,000

About The Position

Prime Intellect is hiring an Applied Researcher focused on computer use to build post-training and evaluation systems for multimodal agents operating browsers, desktops, and cross-application workflows. The work addresses how agents understand changing interfaces, maintain intent over long workflows, recover from mistakes, manage permissions and side effects, and determine whether tasks are complete. The research is grounded in an active external program and focuses on turning real workflows and traces into environments, rewards, and experiments that produce reliable improvement. Applied Research work turns workflows into tasks, environments, evaluations, and reward signals; runs post-training experiments against real baselines; and deploys effective results. Researchers investigate whether gains come from the model, data, harness, inference, or infrastructure. Reusable work may become shared research, open-source systems, and product capabilities. Every role is externally facing, and most researchers work directly with customers, users, or research partners while maintaining a substantive research agenda.

Requirements

Relevant experience in multimodal agents, computer use, browser or desktop automation, reinforcement learning, imitation learning, human-computer interaction, embodied agents, or long-horizon agent evaluation.
Hands-on experience building and debugging model-training, post-training, agent, evaluation, or closely adjacent experimental loops.
Strong experimental judgment, including meaningful baselines, careful trace inspection, clear failure attribution, and willingness to change direction based on evidence.
Strong Python skills and the ability to build artifacts other researchers and engineers can run, inspect, and extend.
Comfort owning ambiguous problems and keeping partners and teammates informed while resolving them.

Nice To Haves

Publications, open-source systems, benchmarks, or shipped products that demonstrate exceptional work.
Ability to describe a difficult project you built, what failed, and how evidence changed your approach.

Responsibilities

Build long-horizon computer-use environments from real workflows while preventing shortcuts, leakage, and unsafe side effects.
Train agents to recover from interface changes, partial failures, and ambiguous intermediate states rather than restarting or silently drifting.
Develop evaluations that distinguish genuine completion from plausible-looking behavior, including correctness, efficiency, permissions, and appropriate escalation.
Build multimodal environments from partner workflows, demonstrations, tools, interface state, and outcome criteria.
Develop browser, desktop, and cross-application agents with reliable sandbox and boundary enforcement.
Train and evaluate agents using demonstrations, interventions, production failures, synthetic tasks, and verifiable outcomes.
Run experiments on recovery and generalization across applications, layouts, permissions, and deployment contexts.
Work directly with customers, users, or research partners while maintaining a substantive research agenda.
Build and inspect meaningful baselines, analyze traces, attribute failures, and change direction when evidence warrants it.