Yohei Nakajima

Yohei Nakajima is the co-founder and managing partner of Untapped Capital and creator of BabyAGI, an early open-source experiment in autonomous task-planning agents. His work combines early-stage venture investing with public experiments in agent memory, coding systems, and controlled self-improvement.

From startup investing to building agents

Nakajima studied economics, finance, and leadership at Claremont McKenna College and began organizing startup communities in Los Angeles in 2009. At Techstars, he helped launch the Disney Accelerator and became director of pipeline, supporting startup sourcing across its accelerator network. He subsequently joined Scrum Ventures as a venture partner and helped lead its corporate-innovation work with companies including Nintendo, Panasonic, and Dentsu.

In 2020, Nakajima co-founded Untapped Capital with Jessica Jackley to invest in founders outside established venture networks. His build-in-public investment strategy uses hands-on technical experiments to understand emerging categories and meet founders. Announcing the firm’s second fund, he connected that approach with proactive sourcing and earlier identification of promising technical shifts.

Projects and technical convictions

• BabyAGI and autonomous task planning. Released in March 2023, BabyAGI gave language-model agents an objective, generated and executed tasks, and created follow-up work from the results. Nakajima treated its early limitations as an invitation to investigate persistent execution history, reusable functions, and agents that can extend their own capabilities.

• Graph-structured agent memory. InstaGraph turns text or a URL into a visual knowledge graph; related experiments include MindGraph and prettygraph. These projects explore how agents can retain connections among evidence, claims, tasks, and decisions instead of flattening everything into conversational text.

• Self-building coding agents. Ditto experiments with generating multi-file applications from natural-language instructions. It reflects Nakajima’s interest in finding the smallest useful systems that can create tools and expand their own capabilities.

• ActiveGraph and event-sourced agents. ActiveGraph organizes agent activity around an immutable event log that records actions and changes to the agent itself. Its underlying architecture projects shared graph state from that history, allowing independent behaviors to coordinate while supporting replay, rollback, branching, and auditable execution.

• Policy-gated self-improvement. ActiveGraph restricts sensitive modifications through explicit policies, human approval, and experimental checks. In his Regimes experiment, proposed changes are evaluated against held-out examples before acceptance. Nakajima’s AI Engineer presentation extends the idea to research agents, coding workflows, and simulations that preserve failed experiments alongside successful ones.