This past weekend at Stanford's TreeHacks 2026, we built Project Vend, a fully agent-driven food bank run by Anthropic's Claude!

Our system included a robotic arm that sorted over 100+ lbs of donations from fellow builders with computer vision, and a series of Claude agents that managed the operation and communicated with Palo Alto shelters, reasoning over resource management and scheduling distribution. 

We placed in Anthropic's "Human Flourishing" track and delivered the food the next day!

When Anthropic released Project Vend showing Claude could run a business autonomously,  we decided to prove that if AI could help people take, it could also help them give!

Beyond the complexities of syncing multi-threaded agents and integrating vision-language models for the robotic arm, we found it to be a great exercise in applied EA. 

It was fun to watch our agents messaging shelters (and correcting each other's occasionally hallucinated promises of food) - proving that the theories about scalable oversight and AI safety are certainly effective when applied to the messy, real world.

And we were proud to show that AI and robotics, when seen through this lens, change from mechanic tools into powerful, amplifying agents of human kindness.

We're excited to keep building in this direction, and hope this inspires others to do the same :)