Yash Pandya

关于

My research focuses on training agents to use tools and computers, learn from experience, and handle failures safely. My work includes the Fara (opens in new tab) computer-use models, Echoverse (opens in new tab) training environments, agentic reasoning (opens in new tab), on-policy learning, and agentic safety (opens in new tab). Lately, I’ve been focused on recursive self-improvement of the entire agentic stack, where models, environments, harnesses, tasks, skills, and memory evolve together. I’m interested in understanding which component needs to change when an agent fails, and whether the changes generalize to improve performance on new tasks. My broader interests include AI agents, post-training, reinforcement learning, continual learning, computer-use agents, on-policy distillation, agentic safety, and agent evaluation.