Chengxi Taylor is co-founder and president of General Reasoning, a London-based AI research company developing agents that pursue complex objectives over extended periods. Her work confronts a central limitation of contemporary AI: maintaining useful memory, judgment and adaptability when decisions unfold across months and their consequences emerge slowly.
Before entering AI research, Taylor led companies developing tools for creative communities. Under her earlier name Chengxi Wang, she became chief executive of the 3D-design marketplace MyMiniFactory in 2019, then founded and led Satori, which developed professional 3D printers for dental and creative applications. Her subsequent entrepreneurial work included augmented reality. She holds an Oxford MBA and the CFA qualification, invests in AI companies, and oversees investment and strategic partnerships at General Reasoning, including deals with Meta and Google DeepMind.
Building agents for the long haul
- Long-horizon reinforcement learning: Taylor studies how agents can continue reasoning when a task exceeds their context window. Her approach combines context compaction and external memory—summarizing prior work, recording intermediate reasoning and searching earlier trajectories—with value models that help address sparse rewards, variable task lengths and difficult credit assignment. Archives must support fresh reasoning without letting agents shortcut evaluations by retrieving previous answers.
- KellyBench: Taylor co-authored this research benchmark, which challenges agents to build predictive models, manage a bankroll and navigate a simulated Premier League betting market across an entire season. Every frontier model evaluated lost money on average, illustrating how conventional short-form benchmarks can conceal weaknesses in uncertainty management and sequential decision-making.
- OpenReward and training tradeoffs: General Reasoning’s OpenReward hosts reinforcement-learning environments, including KellyBench. Taylor emphasizes that pipelined training can keep computing hardware productive but introduces stale data, while value-model bootstrapping generates earlier learning signals at the risk of bias. She has also advocated interactive training environments that expose models to feedback and extended experience.
Taylor also founded SEEN, an interdisciplinary community connecting technologists, artists and other creative thinkers.
