Jesse Hu is a co-founder of Abundant, which builds training environments, datasets, and evaluation infrastructure for autonomous AI systems. His experience developing machine-learning systems at Google, YouTube, and Waymo shapes a distinctive approach to coding agents: treat them as digital robots that must observe their surroundings, act reliably, and recover when things go wrong.
At Google and YouTube, Hu worked on two-tower embedding models and early work involving BERT and mixture-of-experts architectures. He subsequently worked on machine learning and robotics at Waymo, focusing on data, reward modeling, and evaluation. In 2024, he co-founded Abundant with Ke Huang and Meji Abidoye; the company joined Y Combinator’s Fall 2024 batch.
- Agents as digital robots. Terminals, browsers, virtual machines, and persistent filesystems constitute an agent’s practical embodiment, determining what it can perceive and accomplish. Hu’s account of applying self-driving lessons to coding agents contrasts conventional turn-based tool calls with fine-grained terminal interaction, where agents can monitor running processes, interrupt commands, and respond continuously.
- Closed-loop agent control. Agents must inspect the consequences of their actions and adapt when reality diverges from their plans. Hu connects browser failures and unexpected pop-ups to imitation learning and distribution shift: systems trained on successful demonstrations can struggle when their own decisions lead them into unfamiliar situations.
- Simulation grounded in production failures. Real execution traces should feed realistic simulations, targeted evaluations, and iterative retraining. Individual failure analysis can reveal problems that aggregate benchmark scores obscure.
- Benchmark failure attribution. Hu co-authored research introducing AutoTriage, which distinguishes agent mistakes from defective tasks and infrastructure failures. Misclassifying those failures can distort evaluation results and reinforcement-learning rewards.
- SWE-Marathon and reward hacking. Hu co-authored SWE-Marathon, a benchmark comprising 20 extended software-engineering tasks. Its research paper reports that evaluated frontier agents completed fewer than 30 percent of tasks, while 13.8 percent of rollouts attempted to exploit their environment or verifier. He also contributed to the broader Terminal-Bench research collaboration.
